# Vasudev Siddh (Vasu-Devs): AI engineer > I'm Vasu, an AI and full-stack engineer based in India. I build voice tools, desktop apps, and software for clients. Currently an AI Engineering intern at withlayer.ai. Outside work, I'm building JustHireMe, Svara, and Dreamer, and sharing what I learn through code and videos. Open to work: AI engineering roles and freelance builds. Based in India (IST), works your hours. - Site: https://www.siddhvasudev.com/ - Email: siddhvasudev1402@gmail.com - Book a call: https://cal.com/vasu-devs - GitHub: https://github.com/vasu-devs - X: https://x.com/vasu_devs - LinkedIn: https://www.linkedin.com/in/vasu-devs/ - YouTube: https://www.youtube.com/@Vasu-Devs - Instagram: https://www.instagram.com/vasu_._devs/ ## Experience ### withlayer.ai: AI Engineering Intern (June 2026 - Present) I work across the whole stack of a real-time voice AI product: the agent runtime and its truthfulness, security and authorization, streaming voice, a native macOS app, and the background systems behind them. Agent runtime and truthfulness: background and interactive agent turns run on one runtime with a server-controlled final step. Every list result says what it covered and whether it was truncated, every endpoint is classified by what an empty result means, and the honesty rules live in a platform layer no persona can remove. When a small settings tool exposed a runaway loop on live calls, I fixed it in the shared core: a tool the permission layer refuses is no longer advertised, and unsolicited turns have an iteration ceiling. Security and authorization: a default-deny authorization rail where broader reads must be offered and accepted first, bound to the exact query; closed cross-tenant paths, a path-traversal bug and trusted model-supplied identifiers; loggers that cannot accept a raw error, so secrets and personal data can't reach logs. Real-time voice: a streaming speech pipeline with interruption-robust turn detection, a per-turn latency waterfall, and practice calls that ring right after a meeting, with a pre-voiced opening so the agent speaks moments after pick-up. Native macOS app (Swift): signal-level audio fixes, on-device active-speaker detection where no image leaves the Mac, speaker naming that leaves a line unnamed rather than guess, and a release that survives backend outages, with single-flight sign-in, onboarding and settings. Background systems and product: governed scheduled routines with exactly-once publication, event-driven meeting preparation, a today-only home surface where each item has one owner, and approval-gated CRM write-back on every surface. Evals and practice: deterministic scorer gates, LLM-as-judge rubrics, mutation testing, real-Postgres proofs and authenticated end-to-end tests, with what was not verified stated in every change. - Unified background and interactive agent turns on one runtime, so automated runs follow the same rules and audit trail as chat - Built evidence contracts across the agent's API so it never treats "couldn't read" as "nothing there" - Replaced prompt instructions with enforced output contracts, and deleted mechanisms that measured harmful - Built a default-deny authorization rail and closed cross-tenant, path-traversal and trusted-identifier paths - Built the streaming speech pipeline and a per-turn latency waterfall, then optimised against it - Fixed the macOS app's audio pipeline at the signal level and built on-device active-speaker detection - Took the native macOS app to release: outage buffering, single-flight sign-in, onboarding and settings - Built practice calls that ring right after a meeting: an AI roleplay with the same buyer and a coach - Built a today-only home surface where every item has one owner and nothing is sent without approval - Fixed a runaway agent loop in the shared core: never advertise a tool the permission layer refuses - Built governed background execution, event-driven preparation and human-in-the-loop CRM write-back - Built eval suites with deterministic scorer gates, LLM-as-judge rubrics, mutation testing and real-Postgres proofs Stack: Python, TypeScript, React, Next.js, PostgreSQL, WebRTC, LiveKit, Deepgram, Voice AI, LLM Evaluation, Vitest, Playwright, Swift, macOS Development, Supabase, ScreenCaptureKit, AI Agent Development, Authorization ### Confidential client: Full-Stack AI Engineer (Contract) (May 2026 - Present) I run engineering end to end for a venture-backed consumer-AI startup: the AI assistant backend, data acquisition and enrichment, retrieval and recommendations, AWS operations, CI/CD and iOS releases. AI assistant backend: I rebuilt the assistant as a Python/FastAPI service with tool orchestration, retrieval, conversation context and structured responses the mobile app renders, then took it to production. Time to first results went from 72-122 s to under 10 s by profiling every stage, with request deadlines, concurrency and deterministic fast paths. Golden and adversarial evaluation suites came next, because passing unit tests didn't prove useful behaviour. Data acquisition and enrichment: the core acquisition engine (source adapters, pagination, normalisation, deduplication and data-quality gates), and enrichment pipelines that combine deterministic checks, model classification and image and text embeddings, taking coverage from 0% to 99.5%. An uncertain paid request is never silently repeated. Retrieval, search and recommendations: a multi-million-vector retrieval index answering in ~15-21 ms, keyword search from 10.6 s to 53 ms, and a recommender with a closed feedback loop across the AI service, the backend and the iOS app. AWS, CI/CD and releases: ECS/Fargate, ECR, Batch, Step Functions, Lambda, S3, DynamoDB, CloudFormation, IAM and CloudWatch, with guarded rollouts, immutable image identity, cost controls and incident containment. I own CI/CD in GitHub Actions and run iOS release engineering: signing and provisioning, self-hosted Mac runners, TestFlight and App Store submission. Backend and ownership: Node.js/TypeScript work on price and currency consistency, a notification pipeline rebuilt from 100% failed deliveries with truthful reporting, and idempotent credit accounting; onboarding packages, technical handoffs, a security assessment, and incidents owned end to end, including ones I caused. - Rebuilt the AI assistant as a Python/FastAPI service with tool orchestration and retrieval, and took it to production - Cut time to first results from 72-122 s to under 10 s by profiling every stage - Built golden and adversarial evaluation suites, because passing unit tests didn't prove useful behaviour - Built the data-acquisition engine and enrichment pipelines (0% to 99.5% coverage), where an uncertain paid request is never silently repeated - Built a multi-million-vector retrieval index answering in ~15-21 ms; keyword search from 10.6 s to 53 ms - Built a recommender with a closed feedback loop across the AI service, backend and iOS app - Ran AWS operations with guarded rollouts, immutable image identity, cost controls and incident containment - Run iOS release engineering: signing, self-hosted Mac runners, TestFlight and App Store submission - Rebuilt a notification pipeline that was failing 100% of deliveries, with truthful delivery reporting - Review, integrate and release teammates' work, and own incidents end to end, including ones I caused Stack: Python, FastAPI, MongoDB, Vector Search, PyTorch, Computer Vision, ONNX, Redis, Web Scraping, Data Pipelines, LLM Evaluation, Docker, GPU Optimization, Node.js, TypeScript, AWS, GitHub Actions, Swift, TestFlight, Recommender Systems ### Confidential client: Full-Stack Engineer (Freelance) (March 2026 - April 2026) Sole engineer on a production multi-currency profit-and-loss platform, delivered in 18 days. The business was reconciling finances by hand across several payment processors, an advertising platform and a fulfilment provider. Nobody could answer what an individual product actually earned, and because it is financial data any answer needed an audit trail. I built five live integrations behind a single connector interface with encrypted credentials at rest, and a profit-and-loss engine handling refunds, chargebacks, processor fees, deduplication across processors and multi-currency arithmetic in arbitrary-precision decimals rather than floats. Two distinct order flows had to be normalised, including matching upsell charges to their parent orders within a time window - getting that wrong either double-counts revenue or orphans it. A three-phase sync pipeline uses signed continuation tokens so multi-year historical imports survive serverless execution limits. Performance work was measured before and after throughout: 10x faster ingestion, 100x faster currency backfill, 4,500x faster attribution, and a 95%+ reduction in dashboard load time. Data quality was a substantial part of the work. I detected a six-figure double-count, and reverted one of my own in-flight fixes after establishing it would have erased a much larger block of legitimate revenue - the revert was the correct call and mattered more than the original change. I also recovered 641 orphaned orders and 2,350 mis-attributed transactions. Shipped with 85 tests including equivalence tests proving the fast and slow query paths agree, three-tier role-based access control, rate limiting and audit logging. Signed off with zero post-handover bug reports. - Delivered a production financial platform to a European e-commerce group in 18 days as sole engineer, at full scope - Built five live integrations across payment processors, an advertising platform and a fulfilment provider behind a single connector interface - Implemented a profit-and-loss engine handling refunds, chargebacks, fees, cross-processor deduplication and multi-currency arithmetic using arbitrary-precision decimals - Normalised two distinct order flows including time-window matching of upsell charges to parent orders - Delivered 10x faster ingestion, 100x faster currency backfill and 4,500x faster product attribution - Reduced dashboard load time by over 95% using composite indexes, database-side functions and a pre-aggregated rollup table - Detected a six-figure revenue double-count, and reverted my own in-flight fix after proving it would have erased a larger block of legitimate revenue - Recovered 641 orphaned orders and 2,350 mis-attributed transactions - Shipped with 85 tests, role-based access control, encrypted credentials at rest, rate limiting and audit logging - Client signed off with zero post-handover bug reports Stack: Next.js, TypeScript, PostgreSQL, Prisma, Payment Integration, Data Engineering, Performance Optimization, RBAC, Encryption, Vercel ## How I work - **Measure before optimizing.** A latency waterfall before a speed-up, an eval gate before a prompt ships. - **Mechanisms, not instructions.** LLM reliability comes from contracts, reason codes and tests, not from asking the model nicely. - **A regression test for every bug.** Reproduce it first. A fix without a test is unfinished. - **Ship behind flags.** Default off, one-variable rollback, no redeploy. - **Say what wasn’t verified.** If my own experiment loses, I report that it lost. - **Ship the whole slice.** Packaging, CI, signed releases and docs. A demo is not a product. ## Open source 8 merged PRs across 5 external repositories - [kana-dojo](https://github.com/lingdojo/kana-dojo) (4 PRs merged): Authored four themes plus design documentation. - [free-programming-books](https://github.com/EbookFoundation/free-programming-books) (275k★ repo): Added a FastAPI cheat sheet to the developer-resource index. - [pqoqubbw/icons](https://github.com/pqoqubbw/icons) (Animated icon): An animated corner-down-left icon with a path-draw animation. ## Projects ### JustHireMe - Local-First AI Job Search App (2,200+ GitHub Stars) AI products · Creator, sole engineer · live A privacy-first desktop app that finds job leads, filters out the noise, explains why each role fits you, and writes tailored resumes and cover letters - all running on your own machine, with no data leaving the device. - **2,272** GitHub stars - **380.6K** launch views Open: https://justhireme.ai #### How it works Runs on your machine: Tauri shell, React workbench, Python FastAPI sidecar. - Job sources → Normalize - Normalize → Quality gate - Quality gate → Local CRM - Local CRM → Fit ranker - Profile graph → Vectors - Vectors → Fit ranker - Fit ranker → Drafts - Drafts → You review #### The detail A Tauri 2 (Rust) desktop shell wraps a React 19 interface and launches a bundled Python FastAPI service. SQLite stores the local pipeline, a Kuzu graph database models your skills, projects and experience so the app can explain fit rather than assert it, and LanceDB holds semantic vectors. A bundled ONNX embedding model means it works offline with no API key. A deterministic quality gate rejects stale, thin, senior-only and low-context postings with visible reasons before any expensive processing runs. 1,926 backend tests at the v1.7 release gate, 110 tagged releases. The launch reached 380K views and took the project from 800 to 2,200+ GitHub stars in two weeks. #### What I did - Designed and built the entire product solo - desktop shell, UI, backend service, data layer and release pipeline - Wrote the Rust/Tauri shell that packages a Python backend as a sidecar process and ships signed auto-updates - Modelled the professional profile as a graph database so role fit is explained through evidence rather than asserted - Built the deterministic quality gate that rejects weak postings with inspectable reasons before any AI cost is incurred - Bundled a keyless local embedding model so the app works offline with zero configuration - Grew the suite from 620 backend tests at the public alpha to 1,926 at the v1.7 release gate, across 110 releases - Ran the open-source launch that reached 380K views and 2,200+ stars #### Stack Python, FastAPI, React, Rust, AI Agent Development, Vector Database, Tauri, SQLite, LanceDB, Kuzu, ONNX, Playwright, LangGraph, Desktop Application, Open Source Case study: https://www.siddhvasudev.com/projects/justhireme --- ### Svara: Private On-Device Voice Dictation for Windows and Linux AI products · Creator, sole engineer · released System-wide dictation for Windows where recognition runs entirely on your own machine - CPU or GPU. Press a hotkey, speak, and the text appears at your cursor in whatever application you are using. Nothing is uploaded. - **54 ms** to first word (Linux, CPU) - **5.56%** word error rate Code: https://github.com/vasu-devs/Svara #### How it works Recognition never leaves the machine. CPU is a first-class path, not a fallback. - Hotkey → Pre-roll buffer - Pre-roll buffer → On-device STT - On-device STT → Cleanup - Cleanup → Your cursor - On-device STT → Your cursor #### The detail Cloud dictation tools are excellent and unacceptable for anything sensitive - they stream everything you say, including client names and credentials read aloud, to someone else's servers. Svara is the local answer: no account, no subscription, works offline. CPU inference is a first-class path rather than a fallback, so it runs on a laptop without a GPU. It is built to survive real use: a pre-roll buffer so the first syllable is never clipped, streaming transcription, and insertion that knows what app it is typing into - terminal-safe handling, editor-aware paste, and a clipboard fallback when Windows privilege boundaries block input. Unfinished dictation is protected with temporary recovery audio, and transcript content is redacted from ordinary logs. Personal vocabulary, local history, optional LLM cleanup through a local model, tray controls, and a packaged release with an update flow that waits for you. A benchmark harness measures latency and accuracy on your own hardware rather than quoting numbers from mine, and regression tests cover cleanup, insertion, configuration and log redaction. A Linux version (a preview release, October 2026) types live while you speak, polishes each sentence in place with a local model, and takes spoken edits such as "scratch that". Its engine was chosen by measurement on the same machine: 54 ms to the first word and 5.56% word error rate, against 290 ms and 8.33% for the previous engine. A short reading set adapts the model to your voice, and the adapted model is kept only if it wins on held-out lines. Python, faster-whisper, Moonshine, CTranslate2, Windows APIs, Wayland, SQLite, Ollama. Open source under AGPL-3.0. #### What I did - Built the whole application - capture, streaming transcription, insertion, tray UI and packaging - Added a pre-roll audio buffer so the first word of a dictation is never lost - Built app-aware insertion: terminal-safe handling, editor-aware paste and a clipboard fallback across Windows privilege boundaries - Made CPU inference a first-class path so it runs on laptops without a GPU - Protected unfinished dictation with temporary recovery audio and redacted transcripts from ordinary logs - Added personal vocabulary, local history and optional LLM cleanup through a local model - Built a benchmark harness that measures latency and accuracy on the user's own hardware - Wrote regression tests covering cleanup, insertion, configuration and log redaction - Ported it to Linux (Wayland, Hyprland): live typing at the cursor, in-place polish, spoken edits, and an engine picked by benchmark (54 ms first word, 5.56% WER) #### Stack Python, Speech Recognition, Machine Learning, Desktop Application, Whisper, CUDA, PyInstaller, Audio Processing, Open Source, Benchmarking, CTranslate2, SQLite, Ollama, Windows APIs, Local AI, Linux, LoRA Fine-Tuning Case study: https://www.siddhvasudev.com/projects/svara --- ### Odeon: Self-Improving Voice Agent Optimization Platform AI products · Creator, sole engineer · shipped Prompt tuning for conversational agents is guesswork. You tweak, you hope, you ship - with no evidence the new version is better and no difficult traffic to test it against. Watch: https://youtu.be/GFdSe4-c_xQ #### How it works Runs cycles until the averages meet every threshold; SQLite keeps each run and every prompt version. - Personas → Simulation - Simulation → LLM judge - LLM judge → All pass? - All pass? → Optimizer - Optimizer → Agent prompt - Agent prompt → Simulation #### The detail Odeon replaces that with a measurable loop. It generates adversarial user personas - the lawyer, the crying student, the stubborn negotiator - and runs full simulated conversations against your agent, so it faces genuinely hard traffic rather than cooperative test scripts. Every conversation is scored on numerical criteria, and the agent must pass all of them to clear a scenario. Allowing a single-metric win is how prompt tuning produces agents that are empathetic and useless. When the agent fails, an optimizer rewrites the system prompt automatically - and renders the change as a git-style red and green diff, because an automated rewrite you cannot inspect is not an improvement you can ship. Every run is streamed live over a websocket and archived for replay and comparison. Python, FastAPI, WebSockets, LangChain, Groq, React, SQLite. #### What I did - Built the full platform - persona generator, simulation runner, scoring engine, optimizer and frontend - Designed adversarial persona generation so agents face genuinely hard traffic - Implemented multi-criteria scoring where all thresholds must pass - Built the meta-agent that rewrites the system prompt on failure - Built the git-style diff view so every automated rewrite is inspectable - Implemented bi-directional websocket streaming of simulations character by character - Built run archiving for replay and comparison #### Stack Python, AI Agent Development, LLM Prompt Engineering, React, FastAPI, WebSockets, LangChain, Groq, SQLite, LLM Evaluation Case study: https://www.siddhvasudev.com/projects/odeon --- ### GPU Kernels on an H200: Five Inference Problems, Faster Than SGLang AI products · Creator, sole engineer · in progress My entry in a public GPU-kernel competition: faster kernels for five parts of one model's inference - attention decode, a whole decode layer, prefill, the dense projections and mixture-of-experts layers with 4-bit weights - each timed on one H200 against SGLang's kernel in the same run. - **1.77×** faster attention decode than SGLang - **316 / 822** competition rank #### How it works No H200 at hand: check runs double as timing probes, so nothing reaches the judge without passing locally. - 5 problems → Lead agent - Lead agent → Kernel file - Lead agent → Reviewers - Reviewers → Kernel file - Kernel file → Local check - Kernel file → SASS check - Local check → H200 judge - SASS check → H200 judge - H200 judge → 1.25–1.77× #### The detail All five beat the baseline: attention decode 1.77x, prefill 1.37x, mixture of experts 1.27x, projections 1.26x and the decode layer 1.25x. Every problem was scored within about 20 hours of the first submission, moving the overall rank from 506 of 667 to 316 of 822. I had no H200, only a laptop GPU of the previous generation, and a daily quota of judge runs. So each submission is one Python file with CUDA compiled at load time and a Triton fallback, check runs double as timing probes, and nothing goes to the judge without passing locally. A lead agent worked overnight on a schedule, with one agent per problem and separate reviewers. Two lessons from the hardware: FP8 mma.sync is emulated on Hopper, so native FP8 maths means wgmma; and the judge's older compiler allocated registers differently, so every kernel is now checked in SASS with three compiler versions. CUDA C++, Hopper wgmma and TMA, CUDA graphs, Triton, PyTorch, FP8 and NVFP4. #### What I did - Wrote CUDA and Triton kernels for five inference problems, all faster than SGLang on the judge (1.25x to 1.77x) - Built a fallback path so a compile failure on the judge never wastes a scored run - Used check runs as benchmarks: variant sweeps timed on the H200 and returned in the log - Ran a lead agent overnight with one agent per problem, separate reviewers and a judge-quota budget - Found that the judge's compiler clobbered a wgmma register and added SASS checks with three compiler versions #### Stack CUDA, Triton, C++, Python, PyTorch, GPU Programming, Performance Optimization, AI Agent Development Case study: https://www.siddhvasudev.com/projects/gpu-kernels --- ### Visual Product Search Across 3.5M Images, Answering in 21ms Client work · Sole engineer · in production Upload a photo of an outfit and get matching products for every item in it separately - top, trousers, dress, bag, shoes, belt, jewellery. It runs entirely on one laptop GPU: no server, no cloud, no per-query cost. - **3.48M** vectors indexed - **~15–21 ms** retrieval #### How it works Measured with the query photo’s own vector hidden, so a product must be found through its other angles. - Outfit photo → Per garment - Per garment → Embedding - Embedding → GPU router - Verifier → 3.48M vectors - 3.48M vectors → GPU router - GPU router → Scoring - Scoring → Matches #### The detail 3,475,255 image vectors across 617,020 products. Two retrieval backends are selected automatically by available GPU memory: exact cosine similarity over the full resident matrix, or a two-stage approach that uses a compressed coarse pass to shortlist candidates and then re-ranks them exactly. The accuracy came from the details: aggregating scores per product with a bonus when a second gallery angle also matches, treating category and colour agreement as a small nudge rather than an override, deduplicating colourways, and removing 34,000 junk vectors such as fabric swatches, logos and size charts. Accuracy is measured honestly - the evaluation hides the query photo's own vector, so a product has to be found through its other angles. An index verifier re-embeds a random sample and proves alignment. The embedding model was chosen by benchmarking two candidates on this specific catalogue rather than trusting a public leaderboard. #### What I did - Built the entire visual search system - embedding pipeline, index, retrieval engine, detection and evaluation - Embedded 3.48M product images using a parallel pipeline with atomic checkpointing - Implemented two retrieval backends selected automatically by available GPU memory - Designed the product-level score aggregation with a multi-hit bonus across gallery angles - Built multi-garment detection so each item in a photo is searched independently - Removed 34,000 junk vectors and fixed a whole class of taxonomy misclassification - Wrote the index verifier that re-embeds a sample and proves row-to-metadata alignment - Benchmarked two embedding models on the real catalogue and chose on evidence rather than reputation #### Stack Machine Learning, Python, PyTorch, Vector Database, Computer Vision, NumPy, SigLIP, Image Processing, GPU Optimization, Benchmarking Case study: https://www.siddhvasudev.com/projects/visual-search --- ### Voice AI Debt-Collection Agent With Full-Duplex SIP Calls AI products · Creator, sole engineer · shipped An outbound voice agent that places real phone calls, negotiates in real time, and can be interrupted mid-sentence like a real conversation. Watch: https://www.youtube.com/watch?v=VsEfOfwh8XM #### How it works Outbound calls over SIP. When a call ends, a 70B model scores risk, legal flags and the outcome. - Caller → LiveKit - LiveKit → Deepgram STT - Deepgram STT → Groq LLM - Groq LLM → Deepgram TTS - Deepgram TTS → LiveKit - Groq LLM → Risk review #### The detail A risk engine analyses each call's transcript as soon as it is saved. It flags regulatory triggers - bankruptcy, legal representation, cease-and-desist - classifies the person's situation, and tags the outcome as a payment promise, refusal or dispute. A dashboard shows each call's transcript with its risk indicators and aggregates recovery rates across calls. Built as three deployable services so the voice agent's event loop never shares a process with the API. Python, FastAPI, LiveKit, Deepgram, Groq, React, Docker. #### What I did - Built the complete system - voice agent, API, risk engine and dashboard - Implemented outbound telephony over SIP with full-duplex interruption - Built the compliance engine that flags regulatory triggers from each call's transcript - Built automatic outcome classification from the conversation itself - Architected it as three separate services so the voice event loop never shares a process with the API - Built the dashboard showing transcripts with risk indicators #### Stack Voice Assistant, Python, Speech Recognition, AI Agent Development, FastAPI, LiveKit, SIP, Groq, Deepgram, Docker, Real-Time Systems Case study: https://www.siddhvasudev.com/projects/vaani --- ### Socratis: AI Voice Interviewer That Sees Every Keystroke AI products · Creator, sole engineer · shipped An AI that conducts live coding interviews over voice while watching the candidate's editor in real time. Code: https://github.com/vasu-devs/Socratis #### The detail Earlier versions hallucinated about code they had never actually received - praising bugs and inventing complexity analysis. The fix was explicit context injection: every edit in the editor pushes a system message into the agent's context, so the model literally sees each change instead of guessing from conversation. The agent is deliberately quiet. It speaks only when it detects a bug, a change of approach, or a stall, and it hints rather than solves. An interviewer that talks constantly is worse than none. Afterwards it runs a second pass and produces a forensic report: strict time and space complexity, line-by-line fixes, communication analysis citing exact transcript quotes, and a check that cross-references what the candidate claimed verbally against what their code actually does. Next.js, Python, LiveKit, Deepgram, Groq, Monaco Editor, Express, MongoDB. #### What I did - Built the voice agent, the coding interface, the backend and the reporting pass - Solved the hallucination problem with explicit editor-context injection on every keystroke - Designed the intervention rules so the agent speaks only on bugs, stalls or approach changes - Built the forensic post-interview report with complexity analysis and line-level fixes - Implemented the claim-versus-code cross-check that catches candidate overclaims #### Stack Voice Assistant, Next.js, Python, AI Agent Development, MongoDB, LiveKit, Deepgram, Monaco Editor, Express, WebRTC Case study: https://www.siddhvasudev.com/projects/socratis-ai-voice-interviewer-that-sees --- ### BranchGPT: Git-Style Branching for LLM Conversations AI products · Creator, sole engineer · live Long AI chats poison their own context. Every tangent and dead end stays in the history forever, burning tokens and diluting relevance, because a linear list is the wrong structure for how people actually think. Open: https://branchgpt.siddhvasudev.com #### How it works Every message is a node in Neon Postgres; a Git-style tree lets you jump to any branch. - Question → Answer - Answer → Follow-up - Follow-up → Merge - Answer → Fork - Fork → Tangent - Tangent → Llama 3.3 summary - Llama 3.3 summary → Merge #### The detail BranchGPT models the conversation as a graph. Every message is a node, and forking from any point creates a parallel branch that inherits the history up to that fork - so you can explore a tangent with full relevant context without polluting the main thread. Merging is explicit and summarised by a language model, so the parent absorbs the insight rather than the entire transcript. The merge logic filters out shared history and appends only what is genuinely new, and the branch survives the merge rather than disappearing - because a branch that vanishes when merged makes people afraid to merge. A visual tree sidebar renders the whole conversation graph for instant navigation. Also shipped as a browser extension that injects the same branching model directly into an existing AI chat interface, storing everything locally. Next.js 16, TypeScript, Drizzle, Neon Postgres, Groq. #### What I did - Designed and built the graph conversation model with branch inheritance - Implemented LLM-summarised merging so the parent absorbs insight rather than transcript - Made merges non-destructive so branches survive, removing the fear of merging - Built the visual conversation tree navigation - Shipped a companion browser extension injecting the same model into an existing chat interface with local-only storage #### Stack Next.js, TypeScript, PostgreSQL, AI Agent Development, Drizzle, Neon, Vercel AI SDK, Chrome Extension, IndexedDB Case study: https://www.siddhvasudev.com/projects/branchgpt --- ### Agentic RAG for PDFs With Tables, Charts and Figures AI products · Sole engineer · shipped Most retrieval systems only handle text. In research papers, medical reports and technical datasheets the answer is often inside a table, a bar chart or a figure caption - so a text-only pipeline either finds nothing or invents a number. Code: https://github.com/vasu-devs/Waldo #### The detail Ingestion uses a high-fidelity document parser for structure, OCR as a fallback for scanned pages, and a vision model that transcribes figures and charts into searchable text before embedding into a vector database. Retrieval is agentic: it routes between answering directly and searching, grades every retrieved passage for relevance, rewrites failed queries and retries within a limit, and automatically pulls in figure transcriptions when the question is visual. Guardrails matter here. Out-of-scope questions are refused rather than answered, and if the vision model's quota runs out the system falls back to the parser's own captions instead of inventing a description. Python, FastAPI, LangGraph, Qdrant, React. #### What I did - Built the full ingestion pipeline including structure parsing, OCR fallback and vision transcription of figures - Implemented the agentic retrieval loop with relevance grading, query rewriting and bounded retries - Added visual-query detection that pulls figure transcriptions in automatically - Designed the refusal behaviour for out-of-scope questions - Built graceful degradation to parser captions when the vision model quota is exhausted - Built the React frontend #### Stack Python, Retrieval-Augmented Generation, LangChain, Vector Database, FastAPI, Qdrant, LangGraph, OCR, Document Processing, Sentence Transformers Case study: https://www.siddhvasudev.com/projects/agentic-rag-for-pdfs-with-tables --- ### AI Legal Assistant for Indian Law With Citation-Backed Answers AI products · Creator, sole engineer · shipped Statutes are dense, heavily cross-referenced and written in English legalese. Generic chatbots invent section numbers, which for a legal question is a safety problem rather than a quality one. Code: https://github.com/vasu-devs/NyaySaathi #### The detail This answers only from verified source documents. Chunking preserves the statutory structure - Article, Section, Part, Chapter - so every retrieved passage is a citable unit rather than a fragment split mid-clause. Retrieval is reference-aware: a hand-built cross-link map expands a query about one provision to the constitutional article it rests on, and results are re-ranked by exact section match and number proximity, because numeric identifiers are exactly where pure semantic similarity is weakest. A 13-language layer detects script, handles romanised Hinglish, retrieves in English against English statutes, and translates the answer back - far more accurate than embedding multilingual queries directly. If the sources do not contain the answer it says so and points to the official source rather than generating a plausible section number. Runs fully locally with an embedded vector database - no server required. #### What I did - Built structure-preserving chunking that keeps statutory units intact - Hand-built the cross-reference map linking statutes to constitutional provisions for query expansion - Implemented reranking by exact section match and numeric proximity - Built the 13-language layer with script detection, romanised input handling and English-pivot retrieval - Designed the refusal behaviour with official-source fallback - Shipped three product surfaces on one backend including per-document isolated collections #### Stack Python, Retrieval-Augmented Generation, Vector Database, FastAPI, React, Qdrant, Sentence Transformers, Gemini, Document Processing, Multilingual NLP Case study: https://www.siddhvasudev.com/projects/ai-legal-assistant-for-indian-law --- ### Dreamer: Agents That Research and Critique Your Ideas Overnight AI products · Creator, sole engineer · pre-release Save an idea, let agents research and argue with it overnight, and wake up to a document you can actually edit. #### The detail Ideas die in notes apps, and long AI chats are impossible to audit later. Dreamer turns research, planning and critique into a durable document where every agent edit is a reviewable proposal. An idea vault with an API, a web client and a mobile client shares one worker system for research, planning, critique and refinement. Agents never overwrite: they propose attributed section changes with version history, and the owner accepts or rejects. An opt-in discovery mode investigates an idea, records evidence for and against, and applies explicit acceptance rules. Work records are PostgreSQL-backed with leases, heartbeats, retries and idempotent scheduling, so a multi-hour job can crash and resume. Provider routing and usage accounting live outside the workflow, so swapping a model does not touch orchestration. Evidence checks and a publication quarantine mean a model saying an idea is good is not treated as proof that it is. Unit and database integration tests cover proposals, workers, migrations and recovery. TypeScript, Node.js, PostgreSQL, React, React Native, Docker, Vitest. #### What I did - Designed the proposal model so agents suggest attributed, versioned section changes instead of overwriting - Built a PostgreSQL-backed worker system with leases, heartbeats, retries and idempotent scheduling - Separated provider routing and usage accounting from orchestration - Added evidence checks and a publication quarantine to the discovery mode - Built the API, web client and mobile client on one shared worker system - Wrote unit and database integration tests for proposals, workers, migrations and recovery #### Stack TypeScript, Node.js, PostgreSQL, React, React Native, Docker, Vitest, AI Agent Development, Multi-Agent Systems, Job Queues, Full-Stack Development Case study: https://www.siddhvasudev.com/projects/dreamer-agents-that-research-and-critique --- ### Multi-Currency P&L Platform Built Solo in 18 Days Client work · Sole engineer · delivered A multi-store e-commerce business was reconciling its finances by hand across several payment processors, an ad platform and a fulfilment provider. Nobody could answer what an individual product actually earned. I built the internal platform that unified it, as the only engineer on the project. - **~4,500×** faster attribution - **641** orphaned orders recovered #### How it works Signed continuation tokens let multi-year imports outlive the 300-second serverless limit. - 2 processors → Sync engine - Store API → Sync engine - Ad platform → Sync engine - Fulfilment → Sync engine - Sync engine → Postgres - Postgres → Daily rollup - Daily rollup → P&L dashboard - Postgres → P&L dashboard #### The detail Five live integrations sit behind one connector interface with encrypted credentials. The profit-and-loss engine handles refunds, chargebacks, processor fees, deduplication across processors, and multi-currency maths using arbitrary-precision decimals rather than floats. A three-phase sync pipeline uses signed continuation tokens so multi-year historical imports survive serverless timeouts. Performance work, measured before and after: 10x faster ingestion, 100x faster currency backfill, 4,500x faster product attribution, and a 95% reduction in dashboard load time using composite indexes, database-side functions and a pre-aggregated rollup table. About 71,000 lines, a 14-model schema, 24 migrations, 85 tests. Delivered on day 18 at full scope with no post-handover bug reports. Next.js 15, TypeScript, PostgreSQL, Prisma, decimal.js, Vercel. #### What I did - Sole engineer end to end - schema, integrations, business logic, dashboard, security and deployment - Designed a 14-model schema with 24 migrations and 40+ indexes - Built five platform integrations behind one connector interface with encrypted credential storage - Implemented the profit-and-loss engine including refunds, chargebacks, fees and cross-processor deduplication - Normalised two different order flows including time-window matching of upsell charges to parent orders - Delivered 10x, 100x and 4,500x speedups on ingestion, currency backfill and attribution respectively - Cut dashboard load by 95% using composite indexes, database functions and a pre-aggregated rollup - Detected a six-figure double-count, and reverted my own in-flight fix after proving it would erase legitimate revenue - Recovered 641 orphaned orders and 2,350 mis-attributed transactions - Wrote 85 tests including equivalence tests proving the fast and slow paths agree #### Stack Next.js, TypeScript, PostgreSQL, API Integration, Full-Stack Development, Prisma, Stripe, PayPal, Shopify, Vercel, Payment Integration, Data Migration, RBAC Case study: https://www.siddhvasudev.com/projects/fintech-pnl --- ### Product Catalog Engine: 485 Brands, 650K+ Products Nightly Client work · Sole engineer · in production An AI shopping assistant is only as good as its catalogue. Buying product feeds is expensive and slow to expand, and writing a separate scraper per brand does not scale past a few dozen. #### The detail I built a self-growing catalogue engine that discovers candidate stores, vets them, onboards them automatically, fingerprints which e-commerce platform each one runs, and fetches through a tiered strategy - browser-free by default using TLS impersonation rather than headless browsers, which is dramatically cheaper and faster. Around 94% of volume flows through a single universal path with no per-brand code, including a workaround for the platform's own page ceiling so large catalogues are not silently truncated. The rest is covered by five platform adapters and a generic sitemap fallback, so it can be pointed at almost any store with no configuration. Every product carries full lifecycle state - first seen, last seen, delisted - plus per-size stock, verified currency, mined attributes and a quality score. A daily digest flags every brand that broke, dropped, recovered or is new, because a scraper fleet without drift alerting degrades invisibly. 485 brands, 650,000+ active products, 600,000+ curated records, refreshed nightly. #### What I did - Built the entire pipeline - discovery, vetting, onboarding, platform detection, fetching, normalisation and curation - Designed the universal platform path that carries 94% of volume with no per-brand code - Worked around the platform's own pagination ceiling so large catalogues are not silently truncated - Chose TLS impersonation over headless browsers, cutting cost and time dramatically - Built five platform adapters plus a generic fallback so it works on almost any store with no configuration - Implemented full lifecycle tracking so delisting is a recorded fact rather than a silent disappearance - Built the daily drift digest that classifies every brand as broken, dropped, recovered or new - Documented the data ceilings honestly in the README rather than hiding them #### Stack Python, Web Scraping, Data Extraction, API Integration, Database Design, Data Pipelines, SQLite, ETL, Data Quality Case study: https://www.siddhvasudev.com/projects/product-catalog-engine-485-brands-650k --- ### AI Shopping Assistant Passing 54/54 Evaluation Gates Client work · AI Engineer · deploy-ready Conversational shopping is not a search problem. A single sentence like 'something for a summer wedding, under budget, and I hate florals' contains an occasion, a price ceiling, a negative style constraint and an implied formality level. Keyword search returns nothing useful, and a naive AI either interrogates the shopper with a dozen questions or recommends products that do not exist. #### The detail I built the assistant backend: one orchestrator composing specialised components for routing, search, outfit building, vision, memory and trends. Each message routes to exactly one action and returns a single structured response the app renders directly. Retrieval is hybrid - keyword matching for explicit brand and category mentions, vector search for intent - over a catalogue enriched offline by a vision pipeline that tags every product image with style, occasion, season, fit and formality. The key product decision was a strict question budget with an act-first decision ladder. Conversational shopping dies if the assistant plays twenty questions, so it makes a best-effort recommendation and states its assumptions instead. 272 unit tests, 54 of 54 deterministic evaluation gates passing, and 57 live QA scenarios against a 12,700 product catalogue. #### What I did - Built the entire conversational backend and its evaluation suite - Designed the orchestrator composing specialised components with single-action routing per turn - Implemented hybrid keyword and vector retrieval over a vision-enriched catalogue - Built the offline vision enrichment pipeline tagging products by style, occasion, season, fit and formality - Designed the question-budget decision ladder that forces action over interrogation - Shipped visual search as a quantised model so production carries no heavy ML framework - Wrote 272 unit tests, 54 deterministic evaluation gates and 57 QA scenarios - Produced the deployment runbooks, architecture docs and API collection #### Stack Python, FastAPI, LLM Prompt Engineering, MongoDB, AI Agent Development, Vector Search, CLIP, ONNX, Redis, LLM Evaluation, Docker Case study: https://www.siddhvasudev.com/projects/ai-shopping-assistant-passing-54-54 --- ### Multi-Tenant School Management SaaS: 24 Modules, 361 Tests Client work · Sole engineer · shipped A per-school SaaS platform covering the full operational surface: academics, admissions, student records, fees, payments, attendance, exams, timetabling, transport, inventory, payroll, leave, accounting ledger, reporting and document generation. #### The detail Tenant isolation is enforced twice - once in the data access layer and again with database row-level security - because a school platform leaking data between tenants is not recoverable. Payments treat the signature-verified webhook as the only source of truth, with idempotent capture so a payment is recorded exactly once regardless of retries or a user closing the tab. All money is integer arithmetic in the smallest currency unit, never floating point. Highlights across the modules: QR-code attendance check-in, timetabling that guarantees no teacher is double-booked, live GPS transport tracking with geofence alerts, and bulk certificate generation. 228 unit tests and 133 end-to-end tests green in CI, with a real object-storage container in the pipeline so file handling is genuinely exercised rather than mocked. NestJS, TypeScript, PostgreSQL, Prisma, Redis, React, React Native. #### What I did - Architected and built the entire monorepo - API, web portal, mobile app and shared packages - Implemented tenant isolation twice, in the data layer and again with database row-level security - Built the payments flow treating the verified webhook as the only source of truth with exactly-once capture - Wrote all 24 feature modules covering the full school operational surface - Built background job processing for invoicing, reminders and absence notifications - Set up CI with a real object-storage container so file handling is genuinely tested - Wrote 228 unit and 133 end-to-end tests - Produced the deployment setup with production containers and reverse proxy #### Stack TypeScript, NestJS, PostgreSQL, React, Full-Stack Development, Prisma, Redis, BullMQ, Razorpay, React Native, OpenAPI, Docker, S3 Case study: https://www.siddhvasudev.com/projects/multi-tenant-school-management-saas-24 --- ### MapMyRepo: Any Codebase as an Interactive Knowledge Graph AI products · Creator, sole engineer · live Understanding an unfamiliar repository means reading a file tree that tells you nothing about relationships, then opening files one at a time. Neither the structure nor the purpose of a codebase is visible from a directory listing. Open: https://mapmyrepo.siddhvasudev.com #### How it works A file explodes into its functions and classes; an optional token lifts GitHub’s rate limit. - GitHub URL → Repo tree - Repo tree → D3 force graph - Repo tree → Gemini - D3 force graph → Ask a node - Gemini → Ask a node #### The detail MapMyRepo turns a repository or an entire profile into an interactive force-directed graph where every file and folder is a node. A force layout rather than a tree is deliberate: a tree shows hierarchy, a force layout shows coupling, which is the more useful signal when assessing architecture. Each node gets an AI-generated architectural summary on demand rather than for the whole repository up front, which is what makes it affordable on a large codebase. A per-node chat panel answers questions with that file's context and its neighbours supplied, so answers are specific rather than generic. Public repositories load with no authentication at all; an optional token raises rate limits and unlocks private repositories. React 19, TypeScript, D3.js, Google Gemini, Vite. #### What I did - Built the repository parser, noise filter and force-directed graph renderer - Chose a force layout over a tree deliberately to surface coupling rather than hierarchy - Implemented on-demand per-node AI summarisation to keep cost viable on large repositories - Built the contextual chat panel supplying node metadata and neighbours - Made public repositories work with no authentication, with an optional token for private access #### Stack React, TypeScript, D3.js, AI Agent Development, Gemini, Graph Visualization, GitHub API Case study: https://www.siddhvasudev.com/projects/mapmyrepo --- ### SDLC Skill System With Learning Memory for AI Coding Agents AI products · Creator, sole engineer · active AI coding agents have predictable failure modes: they assume silently, over-build, edit code they do not understand, and claim success they never verified. Worse, they repeat the same mistakes in every new session because nothing carries over. Code: https://github.com/vasu-devs/Forge #### The detail This is one plugin containing 16 linked lifecycle skills - understand, plan, test-first, debug, review, verify, ship - where each encodes a rule that can actually be checked, and names the rationalisation it forbids so the agent cannot talk itself out of the process. For example: no production code before a failing test exists; after three failed fixes, stop patching and question the architecture; produce a claim-to-proof table before 'done' is permitted. A memory layer logs every file edit and command, distils them into trigger-and-action lessons through a strict quality gate, and injects the highest-confidence ones at session start. The gate is the important part - ungated memory fills with noise within a week. A companion code-graph builds an import map across six languages, surfaces overly-central files, cycles and dead code, and patches itself incrementally so it never goes stale. Node.js with zero runtime dependencies, local ONNX embeddings. #### What I did - Designed and built all 16 lifecycle skills as checkable rules rather than guidance - Made each skill name the rationalisation it forbids so the agent cannot argue its way out - Built the zero-dependency hook core that logs every edit and command - Built the lesson distiller with a strict quality gate to keep memory signal-dense - Built the multi-language import graph with incremental patching so it never goes stale - Extracted the memory subsystem into a standalone plugin once it proved independently useful #### Stack Node.js, AI Agent Development, JavaScript, Machine Learning, ONNX, Static Analysis, Open Source Case study: https://www.siddhvasudev.com/projects/sdlc-skill-system-with-learning-memory --- ### Remote Control an AI Coding IDE From Telegram, Securely Developer tools · Creator, sole engineer · shipped Long AI coding runs finish while you are away from the machine. Checking on them means either leaving a remote desktop session open or forwarding a port to your home network - both worse than the problem. Code: https://github.com/vasu-devs/GhostSync #### The detail This exposes the IDE to a private bot over an automatically created secure tunnel, so there is no port forwarding and no inbound firewall rule. Send a folder path, fire prompts, and receive live screenshots of the IDE from anywhere. For a remote-control tool the security design is the product: a strict user allowlist so a leaked bot token alone grants no control, request rate limiting, screenshots deleted immediately after sending so the chat history is not a recording of your screen, sanitised logs, and credentials stored locally and never uploaded. Shipped as a single executable so it runs for people who do not have Python installed - the difference between a script and a tool. Python, Cloudflare Tunnel, desktop automation, CustomTkinter, PyInstaller. #### What I did - Built the tunnel bridge that requires no port forwarding or firewall changes - Implemented desktop automation driving the IDE with window detection - Designed the security model - user allowlist, rate limiting, ephemeral screenshots, sanitised logs - Built the desktop GUI and packaged it as a single executable for non-developers #### Stack Python, API Integration, Desktop Application, Automation, Cloudflare Tunnel, Desktop Automation, PyInstaller, Telegram API, CustomTkinter Case study: https://www.siddhvasudev.com/projects/remote-control-an-ai-coding-ide --- ### PocketPad: Use Your Phone as a Precision Trackpad, No App Needed Apps & experiments · Creator, sole engineer · released Turn a phone into a precision-style trackpad for Windows, macOS and Linux, with nothing to install on the phone. Open: https://getpocketpad.vercel.app #### The detail A small host runs on the computer and prints a QR code. The phone opens a web page, and every gesture from a laptop touchpad works: tap, drag, two-finger scroll and right click, pinch to zoom, three- and four-finger swipes, and an app switcher that follows your hand and lands on an app when you lift. The phone owns gestures and look-and-feel; the host owns smoothing, validation and input injection. Gesture preferences are read from the PC's own touchpad settings. Three connection modes are supported: router Wi-Fi, the laptop's own hotspot, and USB. Released as one-file builds for three operating systems, with a pip install and a website. The README states the limit plainly: this is input synthesis, not a registered precision touchpad. Python, aiohttp, WebSockets, JavaScript, PyInstaller, GitHub Actions. #### What I did - Built a small host that prints a QR code; the phone opens a web page and becomes the trackpad - Implemented the full precision-touchpad gesture set, including a live app switcher that follows the hand - Read gesture preferences from the PC's own touchpad settings instead of inventing a second set - Added Wi-Fi, hotspot and USB connection modes with a pairing key that changes on every start - Packaged one-file builds for Windows, macOS and Linux with CI for all three - Verified gesture recognition in an emulated phone with Playwright touch events #### Stack Python, aiohttp, WebSockets, JavaScript, Desktop Application, PyInstaller, pytest, CI/CD, Pointer Events Case study: https://www.siddhvasudev.com/projects/pocketpad-use-your-phone-as-a --- ### Fly District: A Fruit-Fly Connectome Simulated Live in the Browser AI products · Contributor: habitat, dashboard and learning, on an upstream project · live A browser world where a fruit fly lives, learns and dies while its whole measured wiring diagram runs as a simulation beside it. Open: https://fly.siddhvasudev.com #### The detail Built on an existing open-source WebGPU fly-brain project by another author, whose history and notice are preserved. My work adds the habitat, the dashboard, learning across lifetimes and a sponsor-marketplace test. All 139,255 neurons of the measured wiring run as a spiking simulation on WebGPU, and an anatomical body runs in MuJoCo. The fly finds food and water, rests, forages and lays eggs. Directly below the world sit every neuron, a live spike raster and the retina. You can give reinforcement, freeze learning, change a synapse, pulse a neuron or disconnect it. The limits are stated on the page: the wiring and body are measured, the dynamics, senses, physiology and inheritance are authored models, and the whole-organism coupling has not been biologically validated. TypeScript, WebGPU, Three.js, MuJoCo, Vite. #### What I did - Built the habitat on top of an existing open-source WebGPU fly-brain project, with the upstream history and notice preserved - Added a world where the fly finds food and water, rests, forages, lays eggs and dies - Paired the world with one continuous dashboard: every neuron, a live spike raster, the retina, connectivity and lineage - Added two online learners and inheritance of what was learned across generations - Moved blocking body simulation into a worker and cut simulation overhead - Documented exactly what is measured and what is an authored model #### Stack TypeScript, WebGPU, Three.js, MuJoCo, Vite, Simulation, Playwright, Open Source Case study: https://www.siddhvasudev.com/projects/fly-district-a-fruit-fly-connectome --- ### Cadence: An Interactive Timeline of 2,872 AI Model Releases Apps & experiments · Creator, sole engineer · live A curated, sourced timeline of 2,872 AI model releases and milestones from 1943 to September 2026, drawn so the acceleration is visible. Open: https://ai-cadence.siddhvasudev.com #### The detail The page shows every release as a zoomable strip, the gap between releases on a log scale, a day-by-day calendar of the last three months and a searchable index. There is no framework: a canvas timeline, hand-built SVG charts and one JSON dataset. Date precision is treated as data. Each record is day, month or year precise, and month and year records are excluded from the daily calendar so they cannot fake a spike on the first of the month. A fixed as-of date anchors every calculation, so the numbers do not drift with the viewer's clock. The page says what it is not: the catalog is selective rather than exhaustive, and recent years are denser partly because releases are better documented. JavaScript, Canvas, SVG, Vite. #### What I did - Curated and sourced 2,872 model releases and milestones from 1943 to September 2026 - Drew a zoomable canvas timeline of every release, with hover and brush - Built hand-drawn SVG charts for the gap between releases and monthly counts - Treated date precision as data so month-precise records cannot fake a daily spike - Wrote a data build that fails loudly on bad input, and unit tests for the statistics - Shipped it with a strict content-security policy and touch-inspectable charts #### Stack JavaScript, Canvas API, SVG, Data Visualization, Vite, HTML, CSS, Technical Research Case study: https://www.siddhvasudev.com/projects/cadence-an-interactive-timeline-of-2 --- ### Descent Protocol: Browser Raycaster FPS With an Autoplay Agent Apps & experiments · Creator, sole engineer · live A first-person shooter in the style of the 1993 raycaster era that runs entirely in the browser, plus an agent that plays it. Open: https://doomslopus5-5.siddhvasudev.com #### The detail There is no build step, no dependency and no binary asset. The engine, the maps, the textures, the creature sprites and the sound are all generated in code. It is an original game and contains nothing from any commercial title. The engine does textured walls, perspective floor and ceiling casting, depth-sorted sprites, fog and sliding doors, with an adaptive quality ladder that walks up or down to hold frame rate. There are four weapons, four enemy types and a synthesised score. Press one key and an agent plays the campaign: A* navigation, threat scoring, weapon selection, projectile dodging and stuck recovery, driving the same inputs a person does. The test suite includes a headless run in which the agent completes the game. JavaScript, Canvas, Web Audio. #### What I did - Wrote a raycasting engine with textured walls, floor and ceiling casting, sprites, fog and doors - Generated every map, texture, sprite and sound in code, with no asset files - Built an autoplay agent with A* navigation, threat scoring, weapon selection and dodging - Added an adaptive quality ladder that holds frame rate - Wrote 69 tests, including a headless run where the agent finishes the campaign - Shipped it with no build step and no dependencies #### Stack JavaScript, Canvas API, Game Development, Algorithms, Web Audio, Pathfinding Case study: https://www.siddhvasudev.com/projects/descent-protocol-browser-raycaster-fps-with --- ### Fishslop: A 3D Reef-Keeping Browser Game With No Asset Files Apps & experiments · Creator, sole engineer · live Pilot a small submarine through a sunken ruin, feed the fish, bank the coins they drop and grow the reef. Open: https://vasu-devs.github.io/FishSlop_Opus5.5/ #### The detail Every fish, coral, ruin, texture and sound is generated at runtime from code. No image, model or audio file ships with the game. Fed fish drop coins and grow; coins buy new species and sub parts; reef value unlocks tiers. Every minute or so something invisible circles the basin, and a sonar ping reveals it for a few seconds. The interface follows a written token system: a submersible instrument panel in enamel, brass and green phosphor, with the sonar scope as its one loud element. It plays with keyboard and mouse or touch. TypeScript, Three.js, Vite, Vitest, Playwright. #### What I did - Built a 3D underwater game with a pilotable sub, feeding, coins, a shop and unlock tiers - Generated every fish, coral, ruin, texture and sound at runtime from code - Added a sonar mechanic that reveals an otherwise invisible fish - Wrote a design-token system for the instrument-panel interface - Added touch controls and day and night scenes - Covered it with unit and browser tests #### Stack TypeScript, Three.js, Game Development, Vite, Vitest, Playwright, 3D Graphics, Web Audio Case study: https://www.siddhvasudev.com/projects/fishslop-a-3d-reef-keeping-browser --- ### Run Python or C++ and Watch It Execute in 2D and 3D Developer tools · Creator, sole engineer · live Learning algorithms from static diagrams is a mismatch: the diagram shows a state, but the algorithm is the transitions. Existing visualisers only animate a fixed set of pre-written algorithms, which is useless when you are debugging your own code. Code: https://github.com/vasu-devs/DryRunVisualised #### The detail Here you write your own Python or C++, press run, and step through your execution line by line with full variable inspection - rendered simultaneously as an interactive 2D view and a 3D scene with orbit controls and draggable nodes. Python runs entirely in the browser via WebAssembly, so there is no execution server, no sandbox to secure and no per-run cost. C++ compiles through a public compiler API rather than a custom build farm. Instrumentation is injected rather than requested - you write normal code and the tool adds the tracing. Each executed line emits a uniform snapshot of line, stack, heap and output, and a classifier detects each variable's shape (array, grid, graph, linked list, tree) and routes it to the right renderer. That trace-first design is why two very different languages share one visualiser. 50+ prebuilt algorithms, step forward and backward, adjustable speed. #### What I did - Built the trace-first execution architecture that decouples running code from rendering it - Implemented client-side Python execution in WebAssembly, removing all server and sandbox risk - Integrated remote C++ compilation through a public compiler API - Wrote the instrumentation injection for both languages - Built the variable shape classifier that routes arrays, grids, graphs, trees and lists to the right renderer - Built both the 2D and 3D renderers with step-through controls #### Stack Next.js, TypeScript, Three.js, React, WebAssembly, Pyodide, Monaco Editor, Zustand, Code Instrumentation Case study: https://www.siddhvasudev.com/projects/run-python-or-c-and-watch --- ### Paint Your GitHub Contribution Graph, Entirely In-Browser Developer tools · Creator, sole engineer · live Tools that let you draw on your contribution graph typically want your credentials, call the platform API on your behalf, or fabricate commits under identities you do not control - a security problem for a cosmetic feature. Open: https://git-art-iota.vercel.app #### The detail This builds a real, ordinary Git repository entirely inside the browser using an in-memory virtual filesystem, packages it as a zip, and you push it yourself. No server, no API access, no credentials requested, no fake identities. One deliberate constraint: brightness levels map to a capped number of commits per day, chosen so the result stays inside the platform's normal activity range. Pushing that ceiling higher produces an obviously fake graph. Design tools include a brush and eraser, drop-in images quantised to five intensity bands, text rendered through a bitmap font, 30+ templates, and a time machine targeting any year. Next.js 16, React 19, TypeScript, isomorphic-git, memfs, JSZip. #### What I did - Built real Git repository construction entirely in the browser using a virtual filesystem - Designed the commit-density mapping capped to stay within normal activity ranges - Built the drawing tools, image quantisation and text-to-pixel rendering - Created 30+ templates and a year-targeting time machine - Deliberately required no credentials, no API access and no server #### Stack Next.js, React, TypeScript, Git, isomorphic-git, Git Internals, Browser APIs Case study: https://www.siddhvasudev.com/projects/paint-your-github-contribution-graph-entirely --- ### Multilingual Campus Chatbot - Hackathon Semifinalist, Top 15/900 AI products · Creator, team lead · shipped Campus offices answer the same questions every day - fee deadlines, scholarship forms, timetable changes - often in Hindi or a regional language. The answers already exist in circulars and PDFs, but students want a conversation rather than a document search. Code: https://github.com/vasu-devs/PolySEE #### The detail This answers from institutional documents in Hindi, English and three regional languages, with local embeddings so there is no per-query cost. The feature that makes it deployable is the approval workflow. New answers do not go live unconfirmed: each response shows its confidence score and the source passages it used, an administrator approves or rejects, and only approved answers enter the production set. That creates a fact-checked update loop student volunteers can maintain - which matters when a wrong fee deadline has real consequences. Reaches students through an embeddable website widget plus WhatsApp and Telegram, and escalates to a human when confidence is low. Semifinalist at Smart India Hackathon 2025, top 15 of more than 900 teams. #### What I did - Built the retrieval pipeline over institutional documents with local embeddings - Designed the admin approval workflow that gates answers before they reach production - Implemented confidence scoring with human escalation when uncertain - Added Hindi and three regional languages - Built multi-channel delivery through an embeddable widget and messaging platforms - Led the team to semifinalist placing, top 15 of 900+ teams #### Stack Python, Retrieval-Augmented Generation, LangChain, React, FastAPI, ChromaDB, Ollama, Gemini, Express, MongoDB, Multilingual NLP Case study: https://www.siddhvasudev.com/projects/multilingual-campus-chatbot-hackathon-semifinalist-top --- ### Multi-Agent Static Bug Detector for Specialised C++ AI products · Sole engineer · complete Generic linters miss domain-specific mistakes in specialised C++ because they do not know the API's rules. Those rules live only in a hardware vendor's manual, which no static analyser has read. #### The detail The design is context-first. Before any language model sees the code, a lookup agent queries a documentation server and retrieves the relevant API rules. Only then does detection run - with the manual already in context. That ordering matters more than the choice of model, because it converts a vague 'this looks suspicious' into 'this violates the documented contract for this call'. Detection runs two independent layers - deterministic pattern matching and language-model reasoning - and raises confidence where they agree. Pattern matching alone produces false positives; model reasoning alone produces confident hallucinations. Agreement between two different mechanisms is real signal, and anything below a confidence threshold is dropped. Three model providers are interchangeable so the harness can be evaluated across them without code changes. CSV in, CSV out for batch runs. Python, MCP, LlamaIndex, BGE embeddings. #### What I did - Designed the context-first architecture that retrieves API rules before any model sees code - Built the documentation retrieval server with embedding search and caching - Implemented two-layer detection combining pattern matching with model reasoning - Designed consensus-based confidence scoring with a precision threshold - Made three model providers interchangeable for cross-model evaluation - Built batch CSV processing for evaluation runs #### Stack Python, AI Agent Development, Machine Learning, API Integration, MCP, LlamaIndex, BGE Embeddings, C++, Multi-Agent Systems Case study: https://www.siddhvasudev.com/projects/multi-agent-static-bug-detector-for --- ### AI Financial Forensics for Messy Bank Statement PDFs AI products · Creator, sole engineer · shipped Bank statements are a parsing nightmare - every bank uses a different PDF layout, and mobile payment transactions have their own irregular format. And even once parsed, a table of transactions changes nobody's behaviour. Code: https://github.com/vasu-devs/RupeeRoast #### The detail Extraction uses a layout-aware parser as the primary path after a simpler library failed on multi-column statements, with a fallback for edge cases and validated schemas throughout, so a malformed extraction fails loudly rather than producing silently wrong totals in a financial tool. Two models split the work by their actual characteristics: a fast one categorises every transaction, which is high-volume and low-creativity, and a separate one writes the analysis, which is the opposite. Using one model for both would be either slow or bland. The output is a dashboard plus a genuinely readable spending critique. The featured chart is impulsive versus necessary spending - the one visualisation that actually changes behaviour rather than another category pie chart. Next.js 15, React 19, FastAPI, Recharts, Plotly. #### What I did - Built the extraction pipeline with a layout-aware primary parser and a fallback path - Added validated schemas so malformed extraction fails loudly rather than silently - Split categorisation and narrative generation across two models chosen for their different characteristics - Built the analytics dashboard and the interactive charts - Designed the output around behaviour change rather than another category breakdown #### Stack Python, FastAPI, Next.js, Data Extraction, Machine Learning, PDF Parsing, Groq, Gemini, Recharts, Plotly, Data Visualization Case study: https://www.siddhvasudev.com/projects/ai-financial-forensics-for-messy-bank --- ### Cross-Session Memory System for AI Agents, Fully Local AI products · Creator, sole engineer · active Giving an AI agent memory is easy to do badly. Dumping every past observation into context hoards noise, retrieval by embedding alone misses exact identifiers and error strings, and memory written from tool output is an untrusted input channel that persists. Code: https://github.com/vasu-devs/Forge #### The detail This consolidates rather than accumulates. Every candidate memory is decided against the existing set - add a genuinely new fact, update one that supersedes an old one, delete one that has been invalidated - which is what keeps the store small enough to be useful at session start. Retrieval is a hand-built hybrid of semantic similarity, keyword ranking and entity matching, deliberately without an approximate-nearest-neighbour index: at personal-memory scale that index adds a dependency and a failure mode for no measurable speed gain. That trade-off is documented rather than hidden. Memory carries separate timestamps for when something was true and when it was recorded, so stale facts cannot silently outrank current ones, and importance weighting keeps durable decisions from decaying just because they are old. The store is treated as untrusted input against prompt injection. No telemetry, no network calls, local storage only. #### What I did - Designed the consolidation model that adds, updates or deletes rather than accumulating - Built the hybrid retriever combining semantic, lexical and entity matching by hand - Deliberately omitted an approximate-nearest-neighbour index and documented why at this scale - Implemented the dual-timestamp model separating when a fact was true from when it was recorded - Added importance weighting so durable decisions resist decay - Treated the memory store as untrusted input against prompt injection #### Stack Node.js, AI Agent Development, Machine Learning, Vector Database, BM25, ONNX, Hybrid Search, Open Source Case study: https://www.siddhvasudev.com/projects/cross-session-memory-system-for-ai --- ### OSINT to Behavioural Profile: Multi-Agent Research Pipeline AI products · Creator, sole engineer · shipped Walking into a high-stakes meeting you usually know someone's job title and nothing about how to actually communicate with them. That information exists publicly but assembling it takes hours. Code: https://github.com/vasu-devs/Kyoka #### The detail Three agents run in sequence because each stage has a different failure mode and a different ideal model. A researcher gathers the public footprint across professional and social platforms. An analyst synthesises structured personality frameworks from that raw material. A strategist converts the profile into an actionable brief with recommended approaches, things to avoid, and opening lines. A single prompt attempting all three produces confident, unsourced psychobabble - splitting them is what makes the output traceable to evidence. The interface is deliberately not generic: a charcoal and gold editorial aesthetic with radar charts visualising personality skew, and a chat simulator so you can rehearse the conversation before having it. A single command creates the environment, installs both dependency trees and launches both services - because a two-language project that takes twenty minutes to start does not get used. Python, FastAPI, LangChain, React, Vite. #### What I did - Designed the three-stage pipeline splitting research, synthesis and strategy across different models - Built the OSINT gathering stage across multiple public platforms - Implemented structured personality synthesis from unstructured source material - Built the strategy generation producing actionable briefs - Designed and built the frontend including radar visualisation and a rehearsal chat simulator - Wrote a single-command launcher handling both dependency trees #### Stack Python, LangChain, AI Agent Development, React, FastAPI, Tavily, DeepSeek, Gemini, Web Research, Data Visualization Case study: https://www.siddhvasudev.com/projects/osint-to-behavioural-profile-multi-agent --- ### Job Aggregation Service Across 30+ Sources, Callable by AI Developer tools · Sole engineer · complete Job data is fragmented across dozens of applicant tracking systems, job APIs, community boards and feeds, each with a different schema, rate limit and freshness profile. Anyone building job tooling rebuilds this aggregation layer badly and ends up with duplicates and stale postings. #### The detail One call aggregates 30+ sources into ranked, deduplicated, structured results. The pipeline is built around thin per-source adapters, so adding a source means writing an adapter rather than touching the pipeline - which is what keeps 30+ sources maintainable by one person. Every posting gets a market region classification, a freshness tier, a quality score, and deduplication by canonical URL that persists across runs in a local database with full-text search. Three deployment surfaces share one store: a command-line watcher with adaptive pacing that backs off on errors, an HTTP API, and a tool server so an AI agent can call it directly. The source research is a versioned document grading every endpoint as verified or unverified with field mappings, rather than tribal knowledge in someone's head. Python, httpx, SQLite, FastAPI, Docker. #### What I did - Designed the adapter-per-source pipeline so new sources need no pipeline changes - Built adapters for 10 applicant tracking systems, 9 job APIs and community sources - Implemented region classification, freshness tiering, quality scoring and cross-run deduplication - Built three deployment surfaces over one store including an AI-callable tool server - Wrote adaptive watch-mode pacing with jitter and exponential backoff - Produced a versioned source-research document grading every endpoint #### Stack Python, API Integration, Web Scraping, FastAPI, Database Design, httpx, SQLite, FTS5, MCP, Async Programming, Docker, Rate Limiting Case study: https://www.siddhvasudev.com/projects/job-aggregation-service-across-30-sources --- ### Self-Hostable Bookmark Manager With Hybrid AI Search Client work · Sole engineer · in progress Everything worth saving is scattered across five apps that do not talk to each other. Saved content rots when the source disappears, search is keyword-only so you can never find the thing you half-remember, and every hosted read-later service owns your library. #### The detail Capture completes in under two seconds - the item appears immediately and all enrichment runs asynchronously afterwards. That constraint drove the whole architecture. Enrichment fetches the page, extracts readable content, OCR text and dominant colours, generates AI tags and a summary, creates a semantic embedding, and captures both a screenshot and a full-page archive so the content survives the original going offline. Search combines database full-text search with semantic vector search, fused using reciprocal rank fusion - the pragmatic choice when you have two rankers of unknown relative quality and no tuning data. Multiple AI providers are supported with automatic failover on rate limits or errors, keys encrypted at rest, and a usage dashboard. A browser extension captures from any page, pairing to the account through a flow that never stores the user's password. React, TypeScript, PostgreSQL with vector search, edge functions. #### What I did - Built the sub-two-second capture path with fully asynchronous enrichment - Built the enrichment pipeline covering content extraction, OCR, AI tagging, embedding and archival - Implemented hybrid full-text and semantic search fused with reciprocal rank fusion - Built multi-provider AI support with automatic failover and encrypted key storage - Built the browser extension with password-free device pairing - Implemented full-page archival against link rot #### Stack React, TypeScript, PostgreSQL, Vector Database, Retrieval-Augmented Generation, Supabase, pgvector, Chrome Extension, PKCE, Hybrid Search, Encryption Case study: https://www.siddhvasudev.com/projects/self-hostable-bookmark-manager-with-hybrid --- ### AI Team That Runs Influencer Campaigns End to End Client work · Sole engineer · pre-launch Running an influencer campaign means stitching together five disconnected jobs across spreadsheets, inboxes and agency middlemen - parse the brief, develop a concept, find matching creators, draft outreach, produce a contract. It takes weeks and nobody sees the whole pipeline. #### The detail Five specialist roles are implemented as portable agent skills following an open specification, so each can be versioned and reviewed independently of the app rather than living as a string inside it. An orchestrator routes chat messages to the right skill and streams results to a working canvas rather than a chat log, because campaign artefacts are documents that need editing, not messages that scroll away. Security was treated as first-class from the start: per-tenant row-level isolation, server-only privileged credentials, and creator data pulled from the open web treated as untrusted input against prompt injection - a real threat when your agent reads arbitrary public bios and then takes actions. Also includes generative video previews for campaign concepts. Built for both sides of the marketplace with distinct personas. Next.js 14, TypeScript, PostgreSQL, Claude Agent SDK. #### What I did - Built five specialist roles as portable, independently reviewable agent skills - Built the orchestrator routing chat to the right skill and streaming to an editable canvas - Implemented per-tenant isolation and server-only privileged credentials - Treated externally-sourced creator data as untrusted input against prompt injection - Integrated generative video previews for campaign concepts - Wrote the architecture decision records, pricing analysis and launch checklist #### Stack Next.js, TypeScript, AI Agent Development, PostgreSQL, API Integration, Claude Agent SDK, Supabase, Generative Video, Prompt Injection Defense, Multi-Tenancy Case study: https://www.siddhvasudev.com/projects/ai-team-that-runs-influencer-campaigns --- ### Code-First Video Engine With Perceptual Colour and Quality Gates Developer tools · Creator, sole engineer · active Product demo videos are made in a GUI editor, which means they cannot be version-controlled, diffed, reviewed, regenerated when the product changes, or produced programmatically from real data. #### The detail This treats video as React components with typed, validated inputs, so a video is a function of data and a malformed input fails at the boundary instead of rendering something subtly wrong. Two things make it more than a wrapper. First, a perceptual colour system with contrast checking using the modern accessibility model rather than the older ratio formula, which is known to misjudge legibility on dark backgrounds - so generated palettes are verifiably readable instead of eyeballed. Second, an automated quality gate that rejects the generic template look before rendering. A pipeline can point at a repository and generate a demo video from real repository data, and captions are transcribed locally rather than through a paid API. Remotion, React, TypeScript, Three.js, GSAP, D3. #### What I did - Built a data-driven video engine where videos are typed functions of validated input - Implemented a perceptual colour system with modern contrast checking rather than the legacy ratio formula - Built an automated quality gate rejecting generic template output before render - Built a repository-to-video pipeline generating demos from real project data - Integrated local transcription for captions instead of a paid API #### Stack React, TypeScript, Three.js, Motion Graphics, Remotion, GSAP, OKLCH, APCA, Color Science, D3.js Case study: https://www.siddhvasudev.com/projects/code-first-video-engine-with-perceptual --- ### Real-Time Screen Redaction Filter for Live Streaming AI products · Creator, sole engineer · prototype Live-streaming your screen leaks credentials, tokens, private messages and personal data in the fraction of a second before you notice - and by then it is in the recording. #### The detail This captures the screen continuously, reads the text on it, detects sensitive content, and outputs a redacted feed that streaming software consumes instead of the raw screen. The architecture is dictated entirely by the frame budget: at 30 frames per second every stage shares about 33 milliseconds. That is why it uses a high-performance desktop duplication capture path rather than naive screenshots, a headless computer vision build to avoid pulling GUI dependencies into a pipeline that renders nothing, and why frame-timing benchmarks are part of the test suite - 'is it fast enough' is the only question that matters here. Output is presented as a virtual camera, so it works with any streaming software without needing an integration. Python, OCR, OpenCV, NumPy, virtual camera output. #### What I did - Designed the pipeline around a 33-millisecond frame budget - Chose a high-performance desktop duplication capture path over naive screenshots - Built continuous OCR with sensitive-content detection - Output as a virtual camera so it works with any streaming software without integration - Made frame-timing benchmarks part of the test suite #### Stack Python, Computer Vision, Machine Learning, Automation, OCR, OpenCV, Screen Capture, Virtual Camera, Performance Optimization Case study: https://www.siddhvasudev.com/projects/real-time-screen-redaction-filter-for --- ### Premium School Website With Automated Content Migration Client work · Sole engineer · shipped A school had a dated website with hundreds of images and mandatory compliance PDFs hosted on infrastructure they were leaving. A redesign that loses that content is a compliance problem, and migrating 300+ assets by hand is days of tedium. #### The detail I built the redesign and a purpose-built migration pipeline alongside it: a crawler mapped the live site's structure, and a downloader pulled 308 images and 13 compliance documents automatically. The downloader was written with zero dependencies deliberately - a one-off migration script that needs a virtual environment and a requirements file is a script nobody can re-run in a year. The site itself ports around 20 content sections into a modern framework with a considered visual identity, an intro animation and a full mobile responsiveness pass, and embeds an AI assistant scoped narrowly to a curated knowledge base so it answers admissions questions accurately rather than improvising policy. Deployed to the region closest to its users - a free latency win most deployments leave on the table. Next.js 14, React, TypeScript, Tailwind, Framer Motion. #### What I did - Built the full redesign covering around 20 content sections - Wrote a crawler and a zero-dependency downloader that migrated 308 images and 13 compliance documents - Chose zero dependencies for the migration script so it remains runnable years later - Embedded an AI assistant scoped to a curated knowledge base - Deployed to the region closest to the audience #### Stack Next.js, React, TypeScript, Web Scraping, Web Design, Framer Motion, Groq, Image Optimization Case study: https://www.siddhvasudev.com/projects/premium-school-website-with-automated-content --- ### Personal Portfolio With Themed Realms, Case Studies and a Contact API Apps & experiments · Creator, sole engineer · live My own site, rebuilt from scratch to read like a person rather than a template, and to hold up as a working piece of software rather than a static brochure. Open: https://siddhvasudev.com #### The detail The page is generated: one content module holds every project's copy, grouping and status, and a build script renders it into a static page, so facts live in data files and the layout never drifts from them. Two hand-designed themes, a light one and a charcoal dark one, carry artwork at the top and bottom and keep the reading area plain. Every project opens as a case study in place, with its own deep link, focus and scroll restoration, and projects with a recorded walkthrough autoplay it inside the dialog. Project covers are the real application interfaces, not stock art. The contact page posts to a small serverless function that sends mail through a transactional email provider, with a per-IP rate limit backed by a hosted Redis store, and embeds calendar booking. A traffic counter shows views, visitors and who is here now, and a GitHub contribution snapshot is refreshed at build time. Keyboard shortcuts, a printable résumé and reduced-motion handling are built in. Behaviour is checked with Playwright scripts for dialog layout, deep-link context, mobile interactions and both themes before a release. JavaScript, HTML, CSS, Vite, Node.js build scripts, Vercel serverless functions, Playwright. #### What I did - Designed and built the whole site, including both themes and every interaction - Wrote the content module and generator that render every project from data - Built in-place case-study dialogs with deep links, focus and scroll restoration, and autoplaying walkthroughs - Wrote the contact API with transactional email and a per-IP rate limit - Added a live traffic counter, a build-time GitHub contribution snapshot, keyboard shortcuts and a printable résumé - Wrote Playwright checks for dialog layout, deep links, mobile behaviour and both themes #### Stack JavaScript, HTML, CSS, Web Design, Frontend Development, Vite, Node.js, Serverless Functions, Vercel, Redis, Playwright, Accessibility, Git Case study: https://www.siddhvasudev.com/projects/personal-portfolio-with-themed-realms-case --- ### Constraint-Solver Timetable Scheduler for Schools and Colleges Apps & experiments · Creator, sole engineer · shipped Building a school or college timetable by hand is a combinatorial nightmare. Rooms, faculty availability, electives and combined classes all constrain each other, and one change cascades through the whole grid. #### The detail The core of this is a constraint solver rather than a scheduling heuristic. That distinction matters: with a greedy algorithm, 'no teacher is double-booked' is something you test for and hope. Expressed as a constraint, it is guaranteed by construction or the solver tells you the problem is unsatisfiable. I built a standalone solver prototype first, proving the model worked before writing any interface around it. The finished application wraps that engine in three role-based portals - administrators build and publish schedules, faculty see their own load, students see theirs - with authentication and persistence. Python, Google OR-Tools, React, Express, SQLite. #### What I did - Modelled timetabling as a constraint-satisfaction problem rather than a greedy scheduler, which makes conflict-freedom a guarantee rather than a hope - Built a standalone solver prototype first to prove the model was satisfiable before investing in any product around it - Encoded rooms, faculty availability, wings, electives and combined classes as solver constraints - Built the full-stack application with separate Admin, Faculty and Student portals - Implemented authentication, role-based access and persistence #### Stack Python, Google OR-Tools, Constraint Programming, React, Node.js, Express, SQLite, JWT Case study: https://www.siddhvasudev.com/projects/constraint-solver-timetable-scheduler-for-schools --- ### Local Vision-Model Screenshot Organiser, No Cloud Calls AI products · Creator, sole engineer · shipped A folder of four thousand screenshots with no structure. Manual sorting is unrealistic, and uploading four thousand personal screenshots to a cloud vision API is unacceptable. Code: https://github.com/vasu-devs/SSS #### The detail This runs a vision-language model entirely on your own GPU. Each image is understood, classified into one of nine semantic categories, stripped of every URL it contains, and cleaned of formatting noise - producing a per-image log and a master index of every link found. The engineering is in making a vision model usable on consumer hardware: the model loads once at startup, images process one at a time, the GPU cache is cleared on a fixed interval to prevent memory creep, and half-precision weights keep the footprint inside 8GB. That gets it to roughly 5-10 seconds per image on a laptop GPU. No cloud API calls at any point. Python, PyTorch, Qwen2-VL, Transformers. #### What I did - Built a fully local vision pipeline so thousands of personal screenshots never touch a cloud API - Implemented nine-category semantic classification driven by a vision-language model - Added URL extraction and text cleaning per image with a master link index - Tuned it to run within 8GB of VRAM using half-precision weights and periodic cache clearing - Achieved 5-10 seconds per image on consumer hardware #### Stack Python, PyTorch, Qwen2-VL, Transformers, Computer Vision, GPU Optimization, Automation Case study: https://www.siddhvasudev.com/projects/local-vision-model-screenshot-organiser-no --- ### Offline Desktop AI Assistant Running a Local Language Model Apps & experiments · Creator, sole engineer · shipped Asking a quick question to an AI means switching windows, which breaks flow - and cloud assistants need an API key and a connection. Code: https://github.com/vasu-devs/Korosuke #### The detail This is a sidebar that slides in from the screen edge, answers, and slides away. Every response comes from a language model running locally on your own machine, so it works offline with no account and no keys. The engineering point is concurrency. A desktop UI framework blocks its event loop on synchronous inference, freezing the window mid-generation, so generation runs on a background worker thread and streams back to the interface. Without that the app is unusable. Toggling is process-based: running the same shortcut again flips visibility rather than launching a second instance, which avoids needing a background daemon at all. Python, PyQt5, Ollama. #### What I did - Built a sliding sidebar desktop assistant with smooth animation and styled chat rendering - Ran inference on a background thread so the interface never freezes during generation - Implemented process-based hotkey toggling so one shortcut both shows and hides the panel with no daemon - Made it work fully offline against a local model with no API keys #### Stack Python, PyQt5, Ollama, Desktop Application, Concurrency, Local AI Case study: https://www.siddhvasudev.com/projects/offline-desktop-ai-assistant-running-a --- ### AI LeetCode Tutor Extension With Auto-Generated Diagrams AI products · Creator, sole engineer · shipped Getting help on a coding problem usually means leaving the page and pasting into a chatbot that hands you the finished answer, which teaches nothing. Code: https://github.com/vasu-devs/LeetBot #### The detail This injects into the problem page itself with three deliberately different modes: Hint nudges you toward the next step, Tutor explains a concept, and Planner helps you structure an approach. None of them dump a solution. AI explanations are automatically rendered as diagrams, and because a generated diagram in a narrow sidebar is unreadable, they are pannable and zoomable. API keys stay in local extension storage and never transit any server. TypeScript, React, Vite, Chrome Manifest V3, Google Gemini, Mermaid. #### What I did - Built a browser extension that injects directly into an existing site rather than requiring a separate tool - Designed three distinct teaching modes with different pedagogy - hint, explain, and plan - Auto-generated algorithm diagrams from AI output and made them pannable and zoomable in a narrow sidebar - Kept API keys in local extension storage so they never transit a server #### Stack TypeScript, React, Chrome Extension, Gemini, Mermaid, Vite Case study: https://www.siddhvasudev.com/projects/ai-leetcode-tutor-extension-with-auto --- ### Java Desktop Image-to-ASCII Converter, Containerised Apps & experiments · Creator, sole engineer · shipped An image-to-ASCII converter where the engineering is in the responsiveness and the output quality. Code: https://github.com/vasu-devs/PixelForge #### The detail Rendering runs multi-threaded in the background so the interface never blocks, and the quality slider is debounced so a single render fires when you stop dragging rather than dozens as you move. Full-colour mode samples the original image's pixel colours and applies them per character, recreating the image through typography. Export triggers a separate high-resolution render pass at the source image's original dimensions - so a 4K input produces a 4K ASCII image rather than an upscaled screenshot of the preview. The deployment side was the harder problem: packaging a desktop GUI application into a container with a multi-stage build so it can run from a browser tab. Java 17, JavaFX, Maven, Docker, Docker Compose. #### What I did - Built a multi-threaded rendering engine so the interface never blocks during conversion - Implemented slider debouncing so one render fires when the user stops dragging rather than per pixel of movement - Built a full-colour mode sampling original pixel colours per character - Added a dedicated high-resolution export pass at the source image's original dimensions - Containerised a desktop GUI application with a multi-stage build so it runs from a browser tab #### Stack Java, JavaFX, Maven, Docker, Multithreading, Image Processing, CI/CD Case study: https://www.siddhvasudev.com/projects/java-desktop-image-to-ascii-converter --- ### Pathfinding Algorithm Visualiser Comparing Four Approaches Apps & experiments · Creator, sole engineer · shipped Pathfinding algorithms are usually taught one at a time, so their comparative behaviour - which explores more, which guarantees the shortest path, what a heuristic actually buys you - is never visible. Code: https://github.com/vasu-devs/Maze-Pathfinder-Visualizer #### The detail This runs four classic algorithms on the same procedurally generated maze with distinct colour coding, so the difference in how much of the grid each one explores before finding the exit is immediately obvious. Mazes are generated by recursive backtracking, which produces exactly one solution path. Path reconstruction uses parent pointers stored during traversal. The heuristic used for informed search is admissible for this grid, meaning it never overestimates and the shortest path is guaranteed. Runs at 60 frames per second with path length and execution time reported per run. Python, Pygame. #### What I did - Implemented four pathfinding algorithms and a procedural maze generator from scratch - Built side-by-side comparison so the difference in exploration behaviour is visible rather than described - Implemented path reconstruction through parent pointers - Rendered at 60 frames per second with per-algorithm metrics #### Stack Python, Pygame, Algorithms, Data Structures, Data Visualization Case study: https://www.siddhvasudev.com/projects/pathfinding-algorithm-visualiser-comparing-four-approaches --- ### Webcam to Live ASCII Art in the Browser, Nothing Uploaded AI products · Creator, sole engineer · live Real-time video-to-ASCII demos usually upload frames to a server, which is both slow and a privacy problem for a webcam feed. Open: https://ascii-real-time.vercel.app #### The detail This runs entirely in the browser with no build step. Camera frames sample to an offscreen canvas at a dynamic grid size, and each cell's brightness maps to a character from the active palette. Two details make the output good rather than mushy. Brightness uses perceptual luminance weighting rather than averaging the colour channels, so the mapping matches how the eye actually sees. And emoji mode caches each glyph as a bitmap, because emoji have inconsistent advance widths that would otherwise destroy the grid alignment. Optional background removal runs client-side too. All processing stays in-browser - the video never leaves the device. Vanilla JavaScript, Canvas API, MediaPipe. #### What I did - Built real-time frame sampling and character mapping with no build step at all - Used perceptual luminance weighting rather than a naive channel average for accurate brightness mapping - Implemented per-glyph bitmap caching to keep emoji output aligned despite inconsistent character widths - Integrated optional background segmentation running entirely client-side - Kept all processing in-browser so video never leaves the device #### Stack JavaScript, Canvas API, MediaPipe, Computer Vision, Image Processing Case study: https://www.siddhvasudev.com/projects/webcam-to-live-ascii-art-in --- ### AI Email Drafting Tool With Tone and Length Control AI products · Creator, sole engineer · live Writing the same kind of email repeatedly, in the right register, is a real recurring time cost - and generic AI drafts consistently land in the wrong tone. Open: https://email-drafter-three.vercel.app #### The detail Two modes: draft from scratch with a prompt plus optional tone and length hints, or refine an existing draft with instructions. Live preview before you copy or send. Two design decisions worth naming. The draft and refine paths are separate endpoints rather than one endpoint with a mode flag, which gives genuine prompt-level control over each. And every API call in the frontend goes through a single thin wrapper module, so the base URL is environment-swappable and the entire backend can be replaced without touching a single component. Python, FastAPI, React, Vite, Tailwind. #### What I did - Built separate draft and refine endpoints rather than overloading one endpoint with a mode flag - Funnelled all API calls through a single thin wrapper so the backend is swappable without touching components - Implemented tone and length controls at the prompt level - Set up hot reload on both sides for a fast development loop #### Stack Python, FastAPI, React, Vite, API Design, LLM Integration Case study: https://www.siddhvasudev.com/projects/ai-email-drafting-tool-with-tone --- ### Weather App With Structured AI Lifestyle Recommendations AI products · Creator, sole engineer · live A weather forecast tells you the temperature, not what to actually do about it. Open: https://weather-dashboard-ai.vercel.app #### The detail This fuses live weather data with AI-generated clothing, activity and precaution guidance in a single request. The user enters a city, the backend fetches current conditions, passes them to a language model, and returns the merged result. The design decision that makes it reliable is constrained output. The model is instructed to return JSON only, with a fixed shape, so parsing into typed fields is deterministic. Free-form model output would mean regex parsing and silent failures whenever the model decided to be conversational. Python, FastAPI, React 19, Vite, Tailwind. #### What I did - Integrated a live weather API with a language model in a single request path - Used constrained JSON-only prompting so parsing into typed fields is deterministic rather than regex-based - Merged model output with live data into one response shape for the frontend - Deployed both services with a production-ready configuration #### Stack Python, FastAPI, React, API Integration, LLM Integration, Structured Output Case study: https://www.siddhvasudev.com/projects/weather-app-with-structured-ai-lifestyle --- ### Installable Fitness PWA With a Non-Scrolling Viewport Layout Apps & experiments · Creator, sole engineer · shipped Fitness tracking on mobile web is ruined by scroll. The page scrolls, the list scrolls, browser chrome appears and disappears, and the layout shifts under your thumb mid-set. Code: https://github.com/vasu-devs/Health-os #### The detail This locks the entire page to the viewport so it never grows a scrollbar - only the workout and meal lists scroll internally. It uses dynamic viewport units rather than the older fixed unit, which is what makes mobile browser chrome behave instead of causing the classic jump. Installable to the home screen and runs like a native app. The service worker is generated only in production builds, deliberately - a caching service worker in development produces stale-asset bugs that cost hours to diagnose. Next.js, TypeScript, Tailwind, Supabase. #### What I did - Built an installable progressive web app with offline-capable service worker generation - Solved the mobile viewport problem using dynamic viewport units so browser chrome does not shift the layout - Locked the page so only internal panes scroll, eliminating layout shift mid-interaction - Configured service worker generation for production builds only to avoid stale-asset bugs in development #### Stack Next.js, TypeScript, PWA, Service Workers, Supabase, Mobile Web Case study: https://www.siddhvasudev.com/projects/installable-fitness-pwa-with-a-non --- ### Browser Screen Recorder With Floating Camera and PiP Apps & experiments · Creator, sole engineer · shipped Recording a product demo means either installing a heavyweight desktop recorder or using a screen-share tool that gives you no camera overlay and a file you then have to edit. Code: https://github.com/vasu-devs/Ocular #### The detail This runs entirely on browser media APIs - screen capture for the display stream, the media recorder for encoding, and Picture-in-Picture for a floating camera bubble that survives tab switching. Nothing is installed and nothing is uploaded. The recording stays on the machine unless the user chooses to save it. React, TypeScript, Vite, Framer Motion. #### What I did - Built screen and camera capture entirely on browser media APIs with no native install - Implemented a floating camera bubble that survives tab switching using Picture-in-Picture - Built drag and snap behaviour for the camera overlay - Kept recordings local so nothing is uploaded unless the user saves it #### Stack React, TypeScript, Vite, Browser APIs, Media Recording, Framer Motion Case study: https://www.siddhvasudev.com/projects/browser-screen-recorder-with-floating-camera --- ### Cast and Character Overlay for Any Video Site, Spoiler-Safe Apps & experiments · Creator, sole engineer · shipped You pause a show and cannot place an actor. One major streaming platform solves this natively - on itself only. Everywhere else you switch to a search engine and risk spoilers in the results. #### The detail The obvious approach - hover over a face in the video - is impossible, and this project documents why rather than failing at it repeatedly. Content protection makes the video pixels a black rectangle to any software that tries to read them. I verified this in a written research pass before building. What is readable is the page structure, the title and the playback position. So the extension identifies what you are watching and shows who is in it - the same approach the platform-native feature uses, but working everywhere. It works with no API key at all for TV and anime, with an optional free key upgrading it to films and richer data. Spoiler-tagged content is hidden behind a per-section reveal. JavaScript, Chrome Manifest V3, keyless public APIs. #### What I did - Researched the DRM constraint in depth and documented why reading video pixels is impossible, converting a doomed feature into a shippable one - Built title and playback-position detection from the page DOM instead - Integrated two keyless data sources so the extension works with zero setup - Implemented spoiler-safe mode parsing community spoiler tags with per-section reveal - Made it work across any site with a video player rather than a single platform #### Stack JavaScript, Chrome Extension, API Integration, Technical Research, DOM Manipulation Case study: https://www.siddhvasudev.com/projects/cast-and-character-overlay-for-any --- ### AI IPO Research Model With Probability-Weighted Forecasts Apps & experiments · Creator, sole engineer · shipped An investor research tool for a major upcoming listing window, combining a quantitative model with a readable research corpus. Code: https://github.com/vasu-devs/Arohan #### The detail The model has four layers: a regression anchor fitted on a historical dataset of twenty listings spanning 1995 to 2025, a company fundamentals layer, a demand-side overlay, and a macroeconomic gate. Nine scenario cells per company collapse into probability-weighted forecast bands rendered as fan charts, with a live stress slider and bear, base and bull toggles. Computing the model offline and embedding the results as static data means the application needs no backend, no API cost and responds instantly. The honesty is deliberate and built in. The headline scenario would be several times larger than the largest listing in history, so the regression is explicitly labelled an anchor rather than statistics-grade inference, with wide error bars and caveats written into each research file. It is presented as a research exercise, not advice. Built in a single day. React 19, Vite, Recharts. #### What I did - Built a four-layer valuation model from a historical dataset of 20 listings spanning three decades - Computed probability-weighted scenario bands offline and embedded them as static data so the app needs no backend - Built the interactive forecast interface with a live stress slider and scenario toggles - Wrote eight research documents totalling around 4,300 words with 24 cited sources - Labelled the model's limitations explicitly rather than presenting it as more rigorous than it is - Built the whole thing in a single day #### Stack React, TypeScript, Recharts, Financial Modeling, Data Visualization, Technical Writing Case study: https://www.siddhvasudev.com/projects/ai-ipo-research-model-with-probability --- ### LearnAI: An Applied-AI Handbook With Hands-On Playgrounds Apps & experiments · Author · shipped An applied-AI handbook where you poke at tokens, embeddings, retrieval and prompt injection instead of reading about them. Definitions do not show how embeddings or tool use behave; a small toy you can break does. #### The detail It is framed around the question an experienced engineer actually has - what transfers from the skills they already have and what is genuinely new - across language models, retrieval, agents, evaluation, tool protocols, operations and fine-tuning. Every concept has a playground next to the explanation: token sampling, embedding space, retrieval pipelines, prompt-injection awareness, agent failure modes and model-based evaluation. Quizzes, interview practice, and audio and video lessons with transcripts and saved playback position sit alongside. The build decision is deliberate: a standalone docs site with topic navigation, themes and progress tracking, and no build step, so it still opens in five years. The demos are simplified on purpose to expose cause and effect; they are not production tokenizers or retrieval engines. JavaScript, HTML, CSS, Web Audio. #### What I did - Wrote the handbook, framed around what transfers from existing engineering skills versus what is new - Built interactive playgrounds for tokens, embeddings, retrieval, prompt injection, agent failures and evaluation - Added quizzes, interview practice and audio and video lessons with transcripts and saved playback - Kept it a zero-build static site with themes, navigation and progress tracking #### Stack Technical Writing, Documentation, HTML, CSS, JavaScript, Information Architecture, Web Audio, Education Case study: https://www.siddhvasudev.com/projects/learnai-an-applied-ai-handbook-with --- ### AI Readiness Assessment Tool With Scored Feasibility Output Apps & experiments · Creator, sole engineer · shipped Organisations asking whether AI is worth adopting in a given area usually get consultant hand-waving instead of a scored, defensible answer. #### The detail This is a stack-agnostic assessment tool. Around fourteen questions - or an automated scan of a codebase - produce a 0 to 100 feasibility score, a value-versus-feasibility verdict, an implementation catalogue and an indicative cost. The architectural decision that matters is separation: the scoring engine, the scanner and the pricing calculator are pure modules with no DOM access, dual-exported so they can be tested and reused independently of the interface. Scoring logic tangled into UI code cannot be audited, and an assessment tool whose scoring cannot be audited is worthless. Vanilla JavaScript, static HTML and CSS with a custom design system. #### What I did - Built the scoring engine, repository scanner and pricing calculator as pure DOM-free modules - Separated scoring logic from the interface entirely so the scoring is unit-testable and reusable - Designed a 0-100 feasibility score with a value-versus-feasibility verdict - Built an implementation catalogue and pricing output from the assessment result - Designed and built the interface #### Stack JavaScript, HTML, CSS, Software Architecture, Web Design, Business Analysis Case study: https://www.siddhvasudev.com/projects/ai-readiness-assessment-tool-with-scored --- ### Chapter-Based YouTube Comment Sorting Without an AI Model Apps & experiments · Creator, sole engineer · shipped Comments on a long video are an undifferentiated stream. The ones relevant to the section you are actually watching are buried under hundreds about the intro. #### The detail This segments them into chapter-based buckets synced to playback position. The interesting part is what it does not use. An earlier version of this called a language model to classify each comment batch - slow, costly, and it required the user to supply an API key. The current version replaces that entirely with a hand-written TF-IDF and cosine-similarity matcher running in the content script: faster, free, completely private, and good enough for the task. Knowing when not to reach for a model is a skill in itself, and this project is the clearest demonstration of it in my portfolio. Vanilla JavaScript, Chrome Manifest V3, classical information retrieval. #### What I did - Built a browser extension that segments comments into chapter buckets synced to playback position - Wrote a TF-IDF and cosine-similarity matcher by hand, running entirely in the content script - Replaced an earlier LLM-based version after determining classical retrieval was faster, free and private - Shipped both versions and can articulate the trade-off between them #### Stack JavaScript, Chrome Extension, Information Retrieval, TF-IDF, Algorithms, DOM Manipulation Case study: https://www.siddhvasudev.com/projects/chapter-based-youtube-comment-sorting-without --- ### Desktop Streak Tracker With API Integration and Heatmap Apps & experiments · Creator, sole engineer · shipped Streak-based habits fail silently. You notice you broke a ninety-day streak the day after it happened. #### The detail This is a desktop widget that tracks a coding-practice streak through a GraphQL API and reflects its health through a mascot whose mood changes based on whether you are solving inside your usual time window. That is deliberate: an ambient signal in your peripheral vision changes behaviour where a notification gets dismissed. Includes a contribution heatmap built from the historical data. Python, PyQt5, GraphQL, NumPy, Matplotlib. #### What I did - Built a desktop widget consuming a GraphQL API for live streak data - Modelled the user's typical solving time window and surfaced deviation as an ambient signal rather than a notification - Built a contribution heatmap visualisation - Designed the mascot state machine reflecting streak health #### Stack Python, PyQt5, GraphQL, API Integration, Matplotlib, Data Visualization, Desktop Application Case study: https://www.siddhvasudev.com/projects/desktop-streak-tracker-with-api-integration --- ### Browser Automation Bot for Scheduled Data Retrieval Developer tools · Creator, sole engineer · shipped A scheduled automation that retrieves information from a web service that offers no API, and delivers it to a messaging channel. #### The detail Browser automation is often the only option when a service has no programmatic interface, and the engineering is in making it reliable unattended: handling the authentication flow, persisting the session, running on a schedule, and failing in a way that is visible rather than silent. Demonstrates the automation and scraping skill set applied to a real recurring task rather than a demo. Python, Selenium, scheduled execution. #### What I did - Built browser automation to retrieve data from a site with no available API - Implemented scheduled unattended execution - Handled authentication flow and session persistence - Built delivery of the retrieved result to a messaging channel #### Stack Python, Selenium, Browser Automation, Task Scheduling, Web Scraping Case study: https://www.siddhvasudev.com/projects/browser-automation-bot-for-scheduled-data --- ### Local Web-Search Assistant Running Entirely Offline Models AI products · Creator, sole engineer · shipped Quick factual questions usually need a web search plus reading three pages. #### The detail This is a terminal assistant that searches the web for a query and summarises the results using a language model running entirely on your own machine. No query content reaches a model provider. A compact demonstration of combining a search API with local inference through an orchestration framework - the same pattern that scales up to full retrieval systems, at a size where the whole thing is legible. Python, LangChain, Ollama, Tavily Search. #### What I did - Built a terminal assistant that searches the web and summarises results using a locally-run model - Integrated a search API with a local inference runtime through an orchestration framework - Kept all inference local so no query content reaches a model provider #### Stack Python, LangChain, Ollama, Tavily, Local AI, CLI Development Case study: https://www.siddhvasudev.com/projects/local-web-search-assistant-running-entirely --- ### Markdown-Based Operating System for Running a Business With AI AI products · Creator, architect · in progress Business knowledge lives in people's heads and scattered documents. Agent-based automation usually means bespoke code per workflow that nobody maintains, with prompts living nowhere reviewable. #### The detail This separates three things that need different substrates. Directives are standard operating procedures written in prose, because prose is the right format for judgment. Execution is deterministic scripts, because code is the right format for things that must not be improvised. Orchestration reads the context, picks the right procedure and checks the output quality. Everything is plain markdown under version control, which makes the entire operating model auditable, diffable and reviewable - and navigable as a linked graph by both a person and an agent. The organising idea: the model is the employee, the folder is the company. Markdown, git, agent orchestration. #### What I did - Designed a three-layer architecture separating instructions, orchestration and deterministic execution - Chose plain markdown under version control so every process is auditable and diffable - Structured it as a linked knowledge graph navigable by both humans and agents - Wrote the standard operating procedures the agent executes #### Stack AI Agent Development, Technical Writing, Process Design, Git, Automation Case study: https://www.siddhvasudev.com/projects/markdown-based-operating-system-for-running --- ### 3D Avatar AI Companion Running in the Browser AI products · Creator, sole engineer · prototype A conversational AI interface rendered as an animated 3D avatar in the browser rather than a text box. #### The detail Built on a standard avatar format with a web 3D rendering pipeline, with animation driven by conversational state so the character responds visually as well as verbally. Demonstrates real-time 3D rendering and animation in a web context alongside AI integration. Three.js, React, TypeScript, VRM. #### What I did - Built a browser-based animated 3D avatar as a conversational interface - Integrated a standard avatar format with a web 3D rendering pipeline - Implemented animation driven by conversational state #### Stack Three.js, React, TypeScript, 3D Graphics, VRM, Animation Case study: https://www.siddhvasudev.com/projects/3d-avatar-ai-companion-running-in --- ### Advanced Competitive Programming: Tree Query Algorithms in C++ Apps & experiments · Author · ongoing Solutions to contest problems requiring genuinely advanced data structures - the class of problem where a naive approach times out and only a logarithmic-per-query structure passes. Code: https://github.com/vasu-devs/LEETCODE #### The detail The algorithms implemented include virtual trees, binary lifting for lowest common ancestor queries, Euler tours, persistent segment trees and coordinate compression. These are not interview-preparation algorithms; they are what you reach for when a problem has both an offline query structure and a subtree constraint. This matters for client work in a specific way: most performance problems are not solved by faster hardware but by choosing a data structure with better asymptotic behaviour. This is the evidence that I can do that. Also maintain an ongoing solution archive. 5-star problem-solving rating, and selected for elite algorithm training from a cohort of 368. C++ with the standard template library. #### What I did - Implemented virtual trees, binary lifting for lowest common ancestor, Euler tours and persistent segment trees under contest conditions - Applied coordinate compression to make large-range problems tractable - Maintain an ongoing solution archive - Achieved a 5-star problem-solving rating and completed elite algorithm training selected from a cohort of 368 #### Stack C++, Algorithms, Data Structures, Competitive Programming, Problem Solving, Optimization Case study: https://www.siddhvasudev.com/projects/advanced-competitive-programming-tree-query-algorithms --- ### Deep Researcher: Reports Where Every Claim Cites a Verified Quote AI products · Creator, sole engineer · prototype Ask a question, get a report where every claim points at a quote that really exists in a source. #### The detail A report can look well cited while the sources say something else. This engine keeps the evidence attached to each claim from retrieval to the final page, and stops when it cannot back something up. A planner splits the question into bounded workstreams, searches iteratively, and extracts full pages and PDFs. A claim ledger stores each claim with its source, quoted passage, document location and a provenance hash. Checkpoints in SQLite, cancel and restart, live progress events, a REST API, a dashboard and a CLI. Two gates, not one: the quote must exist verbatim in the retrieved text, and the claim must be supported by that quote in context. Passages are picked for relevance to the sub-question instead of truncating documents from the top. When an evidence gate fails, the pipeline stops and keeps the rejected material for inspection rather than papering over it. Tests exercise evidence persistence, fabricated-quote rejection, passage selection and PDF extraction with controlled fixtures. Python, FastAPI, Pydantic, SQLite, HTTPX, pytest. #### What I did - Designed a claim ledger linking every claim to source, quoted passage, location and provenance hash - Built two evidence gates: verbatim quote existence and claim support in context - Built the planner, iterative search and full-page and PDF extraction - Added SQLite checkpoints, cancel and restart, live progress, a REST API, a dashboard and a CLI - Wrote tests for fabricated-quote rejection, passage selection and PDF extraction #### Stack Python, FastAPI, Pydantic, SQLite, HTTPX, pytest, AI Agent Development, Retrieval-Augmented Generation, Hallucination Mitigation, Document Processing Case study: https://www.siddhvasudev.com/projects/deep-researcher-reports-where-every-claim --- ### JustHireMe iOS and Hosted: Job Search With Its Own Backend AI products · Creator, sole engineer · in development The JustHireMe job-search workflow on a phone, with its own account-scoped backend instead of a hole punched into the desktop app. #### The detail The desktop app keeps everything on one machine. A phone needs accounts, sessions, purchase recovery and a backend built for many users. Reusing the local desktop API would cross the wrong trust boundary, so this is a separate build. Email-code sign-in, profile editing, saved opportunities, application tracking, reviewable drafts, sharing, export and account deletion. Native subscriptions and credits through StoreKit, with server-side purchase records and recovery of unfinished transactions. A clearly labelled demo mode with fictional data lets the app be explored without an account or payment. The mobile cloud service scopes everything through authenticated endpoints. Session credentials sit in the Keychain; account content stays in memory over an ephemeral networking session. A purchase is not complete until the server says so. In October the hosted pipeline behind it came together: keyless discovery across the public applicant-tracking systems and job boards (134,362 active jobs from 860 sources), matching with no model call where every match carries its reasons and gaps, guarded auto-apply that never guesses a legal, salary or work-authorisation answer, sign-in with a versioned consent record, payments with signed webhooks, field encryption, backups and an admin console. 1,028 tests pass on PostgreSQL, and a 20-user load test ran with no server errors. A hand-read panel judged 53% of shown matches sensible against an 80% target; the remaining gap is job supply, not ranking. A public beta is planned. Focused on tech roles, a 600-row hand-labelled sample judged 77% of shown matches sensible. Swift, SwiftUI, StoreKit, Python, FastAPI, PostgreSQL, Supabase, Stripe, Alembic, Docker Compose, Caddy. #### What I did - Designed a separate account-scoped mobile backend rather than exposing the desktop sidecar - Built the SwiftUI app: sign-in, profile, saved opportunities, tracking, drafts, export and account deletion - Integrated StoreKit subscriptions and credits with server-side purchase records and recovery - Kept session credentials in the Keychain and account content in memory - Added a labelled demo mode and demo UI tests covering dark mode and larger text - Built the hosted pipeline: keyless discovery over 860 sources, explainable matching, guarded auto-apply, sign-in, payments, encryption and operations - Measured it honestly: a 20-user load test with zero server errors, and a hand-read match panel reported at 53% sensible against an 80% target #### Stack Swift, SwiftUI, StoreKit, iOS Development, Python, FastAPI, PostgreSQL, Supabase, Authentication, Mobile App, Stripe, Web Scraping, Load Testing Case study: https://www.siddhvasudev.com/projects/justhireme-ios --- ### Ori no Michi: Learn Origami From an Interactive 3D Folding Engine Apps & experiments · Creator, sole engineer · personal project Learn origami from a 3D model you can orbit, scrub and pause while you fold along with real paper. #### The detail A flat diagram hides which layer moves and where it should land. I wanted a fold you can look at from any angle, replay and slow down. The engine does rigid polygon clipping and folding with layer selection, mountain and valley motion, unfolding and turn-over steps. The moving layer is chosen from the authored crease segment, so a fold on one flap does not drag unrelated paper along. Previews are generated from the same geometry as playback, so the picture and the animation never disagree. Lessons add crease guides, moving-flap highlights, alignment targets, and a tutorial theatre with camera moves, chapters, captions, narration, speed control and a fold-along pause mode. Reduced motion and keyboard controls are built in. It models rigid folds, not paper physics or collisions. Tests check paper-area conservation, edge lengths, layer selection and chapter coverage. TypeScript, React, Three.js, GSAP, Python. #### What I did - Wrote a rigid folding engine with polygon clipping, layer selection and mountain and valley motion - Chose moving layers from the authored crease so folds never drag unrelated paper - Generated previews from the same geometry as playback - Built the tutorial theatre with camera moves, chapters, captions, narration and fold-along pause - Wrote geometry tests for area conservation, edge lengths and layer selection #### Stack TypeScript, React, Three.js, GSAP, Python, Computational Geometry, 3D Graphics, Animation, Accessibility Case study: https://www.siddhvasudev.com/projects/ori-no-michi-learn-origami-from --- ### EstimateIO: A Browser Game That Trains Your Intuition for Size Apps & experiments · Creator, sole engineer · personal project How big is a blue whale next to a bus? Guess by resizing, measuring or placing things on a scale, then see how far off you were. #### The detail Numbers do not build intuition for size. Dragging an object next to something familiar, and being shown your error at a shared baseline, does. Three modes: resize an object beside a reference, construct a length with your hands, or drop an object on a logarithmic size line. Every measurement is stored in centimetres with explicit conversions and a reference source attached. Scoring is symmetric, so guessing half the size and double the size are equally wrong. Pinch and drag handle pointer add, release, cancel and near-coincident fingers without jumps, and keyboard and buttons work too. Logic tests cover scoring, conversions, log mapping, challenge rotation and gesture transitions. TypeScript, React, Pointer Events, Web Audio. #### What I did - Designed three estimation modes including a logarithmic size line - Stored every measurement in centimetres with explicit conversions and cited references - Made scoring symmetric so over- and under-estimates are judged equally - Built robust multi-touch pinch and drag plus keyboard controls - Wrote logic tests for scoring, conversions, log mapping and gestures #### Stack TypeScript, React, Game Development, Pointer Events, Web Audio, Accessibility Case study: https://www.siddhvasudev.com/projects/estimateio-a-browser-game-that-trains --- ### Reiatsu: Live WebGL Wallpapers Behind Your Windows Desktop Icons Apps & experiments · Creator, sole engineer · prototype A Windows wallpaper studio that runs live wallpapers behind the desktop icons: still images brought to life with WebGL drift, refraction and light, chosen from a floating scene panel. #### The detail A personal experiment in rendering a web scene into the Windows desktop layer without getting in the way of the icons or the work on top. Electron, WebGL, Windows. #### What I did - Rendered a WebGL scene into the Windows desktop layer behind the icons - Wrote drift, refraction and light effects that animate still images - Built a floating scene picker #### Stack Electron, WebGL, Windows APIs, Shaders, Desktop Application Case study: https://www.siddhvasudev.com/projects/reiatsu-live-webgl-wallpapers-behind-your --- ### habiTurtle: Desktop Habit Tracker With a Streak-Driven Mascot Apps & experiments · Creator, sole engineer · shipped A desktop habit tracker with streaks and a turtle that looks happier the more you show up. Code: https://github.com/vasu-devs/habiTurtle #### The detail An ambient mascot in the corner of your screen is a gentler nudge than a notification you swipe away. Python, PyQt. #### What I did - Built the habit and streak tracking - Designed a mascot whose mood follows your consistency #### Stack Python, PyQt5, Desktop Application Case study: https://www.siddhvasudev.com/projects/habiturtle-desktop-habit-tracker-with-a --- ### HoleEmAll: A Small Pygame Arcade Game Apps & experiments · Creator, sole engineer · shipped A tiny arcade loop in Pygame: move, swallow, grow, beat the countdown. Code: https://github.com/vasu-devs/HoleEmAll #### The detail Python, Pygame. #### What I did - Built the game loop, growth mechanic and countdown #### Stack Python, Pygame, Game Development Case study: https://www.siddhvasudev.com/projects/holeemall-a-small-pygame-arcade-game --- ### FounderFinder: 500,000 Startups, Founders and Open Roles, Ranked AI products · Creator, sole engineer · in progress A local dataset and workspace for finding early-stage startups worth working with: 541,695 companies, 570,838 founders and 51,065 open roles, collected only from public, permitted sources. #### The detail A job search aimed at early startups has no single list to search. The companies are spread across accelerator batches, funding news, regulatory filings, company registers in a dozen countries and job boards. This pipeline collects them, merges duplicates, classifies every company with explainable rules and scores each one twice: as a lead, and for fit against a personal profile. Every point in both scores carries its reason. Classification is deterministic and was measured, not assumed: an independent model graded 240 random companies, and the weakest dimension is reported alongside the strongest. A founder email is marked published only when it appears on the company's own site; anything else is stored as a pattern guess and labelled that way. robots.txt and crawl delays are honoured, and sites whose terms forbid scraping are not used. A local website turns the data into a daily workflow: a brief in plain sentences, a 28-facet explorer, company sheets, a pipeline board and an outreach desk. Exports include a formula-driven Excel workbook and a shareable copy with all personal data removed. A model ranks the companies that have enough text for relevance in two passes: a small model screens tens of thousands in batches, and a larger one re-scores the top thousand. Companies that would clash with existing agreements get their own tier instead of being ranked. Python, SQLite, FastAPI, Pydantic, pytest. #### What I did - Built 24 collectors over accelerator portfolios, funding news, SEC filings, nine national company registers and public job boards - Merged the sources into one company record with founders, open roles and evidence signals: 541,695 companies and 570,838 founders - Wrote a deterministic classifier over 25 dimensions and had an independent model grade 240 random companies - Built an explained lead score and a personal fit score where every point carries its reason - Built the local website: a 28-facet explorer, company sheets, a pipeline board and an outreach desk - Kept collection to permitted public sources, labelled every guessed email as a guess, and wrote 267 tests - Ranked 60,864 companies for relevance with a two-pass model screen, with a separate tier for conflicts with existing agreements #### Stack Python, SQLite, FastAPI, Web Scraping, Data Pipelines, Data Extraction, Pydantic, pytest, API Integration, Database Design Case study: https://www.siddhvasudev.com/projects/founderfinder-500-000-startups-founders-and --- ### jevbrowse: Grounded Browser Automation for AI Agents (SDK + MCP) AI products · Creator, sole engineer · prototype A browser runtime that lets any agent or program drive a real browser by intent, without reading raw accessibility trees or generating text for every click. #### The detail Browser agents are slow because a text model reads the whole page for each action, and they fail quietly because nothing proves an action worked. jevbrowse splits the job. The outer agent decides what to do. The runtime takes one indexed snapshot of the live page, asks a small decision-only model where each intent lands among candidates the code supplies, and executes through Playwright on the real DOM node. Every call returns typed evidence: page state, whether the page changed, decision probabilities and timing. An ambiguous target comes back as ambiguous with its alternatives, so the caller chooses instead of the tool guessing. When an intent has exactly one unambiguous match, a local heuristic answers with no network call. Four surfaces share one runtime: a Python SDK, an MCP server over stdio and streamable HTTP, a CLI and an HTTP JSON API. A pytest plugin turns verified flows into re-runnable tests with an on-disk decision cache. Python, Playwright, Pydantic, MCP, FastAPI, pytest. #### What I did - Split browser use into two roles: the outer agent decides what to do, the runtime decides where it lands and whether it worked - Built one-snapshot observation and one decision request per intent, with a local exact-match short-circuit - Returned typed evidence for every call: page state, change detection, decision probabilities and timing - Shipped four surfaces on one runtime: a Python SDK, an MCP server, a CLI and an HTTP API - Added file upload, embedded application-form frames, captcha detection and a human-in-the-loop handoff - Wrote a pytest plugin so verified browser flows become fast, re-runnable tests, plus 423 tests of the runtime itself #### Stack Python, Playwright, MCP, Browser Automation, AI Agent Development, FastAPI, Pydantic, pytest, API Design Case study: https://www.siddhvasudev.com/projects/jevbrowse-grounded-browser-automation-for-ai --- ### AgentChat: Run and Approve Your Coding Agents From Your Phone Developer tools · Creator, sole engineer · prototype Chat with the coding agents running on your laptop from your phone, and approve or deny each command they want to run. #### The detail A long agent task stalls on its first permission prompt the moment you walk away. AgentChat puts each agent session in its own chat, with notifications and Allow once / Deny buttons. There are two ways in: a Telegram bot that takes about three minutes to set up, or a private web app over your own network. The private path is end-to-end encrypted with per-device-pair keys and forward secrecy, so the relay only ever sees ciphertext. Pairing shows the same security code on both screens. Each message shows its delivery state, a failed message can be re-sent, and a device can be revoked. The relay has no runtime dependencies, stores state transactionally, recovers from a crash without manual cleanup and ships as a container with automatic HTTPS. Goals turn it into more than a remote: describe what done looks like, and the agent works on its own branch until every condition is proven again from scratch, then a second agent on the other runtime re-checks the work before it can be called done. Push notifications arrive with the app closed and never carry agent text, and there is an Android app. Each Claude or Codex login is its own agent, and any earlier terminal session can be continued on the phone from a copy, so the original is never touched. An independent security review found eight issues, each fixed with a test. Node.js, SQLite, Web Crypto, Web Push, PWA, Android, Docker, Playwright. #### What I did - Built a relay and a desktop connector that put each coding-agent session in a chat on a phone - Added Allow once / Deny buttons for every command an agent wants to run - Designed per-device-pair end-to-end encryption so the relay only sees ciphertext - Made the relay self-healing with transactional storage, structured logs and a health endpoint - Built a responsive web app with delivery receipts, offline queues and device revocation - Grew the suite from 72 to 171 unit tests, plus browser and stress runs - Built goals with cross-runtime verification: one agent works, an agent on the other runtime must sign off - Added Web Push with the app closed (encryption checked against the RFC test vector) and an Android app #### Stack Node.js, JavaScript, SQLite, End-to-End Encryption, PWA, Telegram Bot API, Docker, Playwright, AI Agent Development, Web Push, Git Case study: https://www.siddhvasudev.com/projects/agentchat-run-and-approve-your-coding --- ### ClientFinder: A Freelance-Client Hunt That Runs Every Four Hours Developer tools · Creator, sole engineer · in progress A private workspace that finds freelance clients every morning and never sends anything on its own. #### The detail Finding contract work means checking a dozen boards a day, most of which show full-time jobs, people offering themselves, or projects already buried under proposals. ClientFinder scans 13 free sources, keeps only people paying for contract work, ranks them against a catalogue of offers, and re-checks that each top pick is still open. An unknown result, such as a rate limit, is never treated as closed. Ranking rewards fresh, uncrowded projects and penalises crowded ones. A scheduled run writes a daily digest with the ranked picks, why each fits, what to verify and a draft for each. Leads move through a pipeline board with notes and follow-up dates. Paid search and model drafting are optional and are never called unless a key is connected and the action is clicked. It now runs every four hours and syncs with a private hosted copy behind a fail-closed login. An outreach desk queues drafts with two follow-ups and a free one-page teardown of the prospect's own site, and once a week it turns my real GitHub activity into proof lines I approve before use. Nothing is ever sent by the app. TypeScript, React, Drizzle, SQLite, Cloudflare Workers, D1, Playwright. #### What I did - Built a local workspace that scans 13 free sources every morning for people paying for contract work - Filtered out full-time postings and self-promotion so top picks are clients only - Ranked for fit, freshness and low competition, and re-checked each top pick live - Added a pipeline board with stages, notes and follow-up dates - Wrote a scheduled daily run that produces a ranked digest with drafts for review - Wrote an end-to-end script that drives the real UI against live sources and reports pass or fail per step #### Stack TypeScript, React, Drizzle, SQLite, Web Scraping, API Integration, Automation, Playwright, Task Scheduling, Cloudflare Workers Case study: https://www.siddhvasudev.com/projects/clientfinder-a-freelance-client-hunt-that --- ### MemOps: Coding-Agent Tasks That Clean Up Their Own Processes Developer tools · Creator, sole engineer · prototype Launch a coding task through MemOps and everything it started stops when it ends. #### The detail Coding agents start dev servers, test runners, browsers and containers, then finish and leave them running. MemOps launches a command inside a Windows Job Object and records what it owns. When the command finishes, fails, times out, or the supervisor itself crashes, the whole process tree stops. An isolated Docker Compose stack can be started and stopped the same way, with data volumes preserved. Command success and cleanup success are reported separately, each with a receipt. A background worker reconciles exactly the resources a task owned after a crash, and never touches live owners or unrelated apps. It ships as a native Windows app with Python bundled, a CLI and a standalone dashboard. It has no runtime Python dependencies. The limits are stated: only commands launched through it are covered. Python, Windows Job Objects, Docker, SQLite. #### What I did - Ran each task inside a Windows Job Object with recorded ownership, so its whole process tree stops with it - Handled completion, failure, timeout and a crash of the supervisor itself - Added owned Docker Compose stacks that stop their containers and keep their volumes - Reported command success and cleanup success separately, with a receipt - Built a background worker that reconciles exactly the resources a task owned after a crash - Shipped a native Windows app, a CLI archive and a standalone searchable dashboard #### Stack Python, Windows APIs, Docker, SQLite, Desktop Application, CLI Development, Process Management Case study: https://www.siddhvasudev.com/projects/memops-coding-agent-tasks-that-clean --- ### MarketMe: From a Product URL to Customer-Intent Signals Apps & experiments · Creator, sole engineer · prototype Give it a product URL. It returns people already describing the problem in public, and the niche agencies that reach them. #### The detail The crawler reads the landing page and up to two linked product pages, derives its own search terms, and searches public sources. Each result keeps its original evidence, a contact route and a suggested conversation starter. The output is researchable opportunities, not confirmed customers, and nothing is ever sent. It runs as a command-line tool with no runtime dependencies, or as an authenticated service with a dashboard: durable jobs, isolated tenants, idempotency and cancellation. One release fixed a real isolation bug, where results from one product could be reused for another. Each run now analyses only the submitted site and records the URL it was asked about. Python, SQLite, Docker. #### What I did - Built a crawler that reads a product page and derives its own search terms - Found public posts describing the problem, and niche agencies that reach those people - Kept the original evidence and a contact route beside every result - Built an authenticated service with durable jobs, isolated tenants, idempotency and cancellation - Fixed an isolation bug where results from one product URL could be reused for another - Kept the discovery CLI free of runtime dependencies #### Stack Python, Web Scraping, API Design, SQLite, Data Extraction, Docker, REST API Case study: https://www.siddhvasudev.com/projects/marketme-from-a-product-url-to --- ### Orchard: Finding iPhone App Niches Where Small Publishers Earn Apps & experiments · Creator, sole engineer · prototype A research dashboard for choosing an app to build from where small publishers already earn, not from an idea. #### The detail It discovers up to the top 100 grossing apps in each of 27 App Store categories, reads public revenue estimates for them, and finds niches: groups of at least two differently named publishers earning above a floor you choose. One scheduled collector runs per category, with locks so a scheduled and a manual scan cannot overwrite each other. Interrupted scans save their completed batches and resume. A failed re-check keeps the previous figure but removes it from niche evidence. The limits are part of the interface: amounts are a provider's rounded estimates rather than verified revenue, coverage is bounded, and an upper bound is never shown as a point estimate. Next.js, React, TypeScript, Drizzle, PostgreSQL, Docker. #### What I did - Built a dashboard that discovers top-grossing apps across 27 App Store categories - Read public revenue estimates and grouped apps into niches with at least two different publishers - Ran one scheduled collector per category with locks so scans cannot overwrite each other - Made slow scans resumable and kept failed re-checks out of the evidence - Kept upper-bound estimates as upper bounds instead of point values - Deployed it privately behind sign-in, with a local Docker mode #### Stack Next.js, React, TypeScript, Drizzle, PostgreSQL, Recharts, Docker, Task Scheduling, Data Extraction Case study: https://www.siddhvasudev.com/projects/orchard-finding-iphone-app-niches-where --- ### CamBud: iPhone Camera App That Shows Groups Where to Stand AI products · Creator, sole engineer · in development A native iPhone camera app for posing: open the camera and step into the shadows. #### The detail The app opens straight into the camera, counts the people in frame with on-device vision, and draws anonymous pose shadows showing where each person should stand. One person gets one shadow; two get a paired arrangement; groups go up to six. Matching never scores attractiveness or infers gender. Detection is stabilised so the guides do not flicker: a change in the number of people must persist before the arrangement changes, and targets stay fixed during the countdown. Only sensor data enters the encoder, so overlays cannot appear in the saved photo. The iOS source is written. Native compilation and real-device verification are still pending, because it was built without access to a Mac. Swift, SwiftUI, AVFoundation, Vision, StoreKit. #### What I did - Designed a camera-first app that counts the people in frame with on-device vision - Drew anonymous pose shadows showing where each of one to six people should stand - Stabilised detection so a count change must persist before the guides change - Kept overlays out of the capture pipeline so the saved photo is clean - Added a local purchase journal that charges only after a successful save - Wrote a browser preview of the interaction, and stated that native compilation is still pending #### Stack Swift, SwiftUI, iOS Development, Computer Vision, StoreKit, XCTest, AVFoundation Case study: https://www.siddhvasudev.com/projects/cambud-iphone-camera-app-that-shows --- ### A Forensic Report on 14 Months of My Own AI Coding-Agent Sessions AI products · Creator, sole engineer · complete A local pipeline that reads every AI coding-assistant session on one machine and turns it into a report on how its owner works. #### The detail Over fourteen months it covers 936 sessions, 6,854 prompts, about 1,274 active hours and 155 projects across eight assistants. The pipeline extracts each session, reconstructs what the agents did, mines the commits in every repository and links sessions to commits. Sessions are then scored against a written rubric with the dates and tool names removed, so a grader cannot judge by era, and sixty sessions are scored twice to report agreement. The findings are specific: a third of all prompts were typed at night, session quality was lowest then, and about one scored session in eight was a loop. A self-check validates invariants and scans for secrets before anything renders. Nothing is uploaded. Python, JavaScript. #### What I did - Read the local session stores of eight coding assistants into one dataset - Linked sessions to the commits they produced across every repository - Scored sessions against a written rubric, blind to date and tool, with 60 scored twice for agreement - Computed behaviour and outcome statistics with confidence intervals - Added a self-check for invariants and leaked secrets before rendering - Built an interactive report, a playbook and a PDF #### Stack Python, Data Analysis, LLM Evaluation, Data Visualization, JavaScript, Technical Writing, Git Case study: https://www.siddhvasudev.com/projects/a-forensic-report-on-14-months --- ### Catalogue Enrichment Pipeline on AWS With Budget-Admitted Runs Client work · Engineer (contract) · in production A product-catalogue pipeline on AWS where quality and cost are part of the processing contract. #### The detail A scraper that works on one laptop is not a data platform. In the cloud it can fail quietly three ways: a job that succeeds and returns nothing, a paid model request repeated after a timeout, and a release nobody can trace to its source. Every run is tied to immutable source artifacts and a qualified worker image, and carries an explicit budget and admission evidence. A record is acquired, then a validated candidate, then a released item, and the three are never counted as one. Unchanged products are skipped, verified work is reused, and an uncertain paid request stays uncertain until reconciled. Styling-attribute coverage on the live catalogue went from 0% to 99.5%. Releases require the exact merged source, an immutable image identity, a clean vulnerability scan and a rollback plan. The record keeps what went wrong as well, including a cutover that was rolled back. Python, AWS Batch, Step Functions, Lambda, ECS, S3, DynamoDB, Docker. #### What I did - Turned a laptop scraper into a governed cloud pipeline: acquisition, classification, embeddings, validated release - Tied every run to immutable source artifacts, a qualified worker image and an explicit budget - Kept uncertain paid model requests in an uncertain state instead of repeating them - Took styling-attribute coverage on a live catalogue from 0% to 99.5% - Built release controls: exact source, immutable image identity, scan gates and rollback plans - Ran cost audits and kept an incident record, including a rolled-back cutover #### Stack Python, AWS, Docker, Data Pipelines, Machine Learning, Embeddings, CI/CD, MongoDB, GitHub Actions Case study: https://www.siddhvasudev.com/projects/catalogue-enrichment-pipeline-on-aws-with --- ### iOS Release Engineering: Signing, TestFlight and App Store Delivery Client work · Engineer (contract) · delivered Carrying a change from reviewed source to a build on a phone, inside an existing mobile product. #### The detail A mobile release is a chain of contracts: backend behaviour, provider delivery, credit accounting, signing identity, upload, store review and device acceptance. Each was verified separately. On the Apple side: a separate development app identity, validation on every merge with uploads only on request, and signing failures diagnosed link by link, from an ambiguous distribution identity to a missing intermediate certificate. Candidate selection, internal TestFlight distribution and an App Store submission followed. On the backend: currency conversion, filtering and sorting made to agree; a notification path that reported success and delivered nothing replaced with one that reports truthfully; and credit accounting made safe to replay. When a later candidate could not pass its UI test run on the build host, it was reported as blocked and not uploaded. Swift, Xcode, TestFlight, Node.js, TypeScript, MongoDB, GitHub Actions. #### What I did - Set up a separate development app identity so internal testing can sit beside the customer app - Separated validation on every merge from explicitly requested uploads - Diagnosed signing failures across identities, certificate chains, profiles and archive settings - Managed candidate selection, TestFlight distribution and an App Store submission - Fixed backend contracts the app depended on: currency consistency, notifications and credit accounting - Reported a blocked release candidate as blocked instead of weakening the release gate - Fixed two self-hosted runner races found by failed release attempts, then shipped the next public release #### Stack Swift, iOS Development, Xcode, TestFlight, CI/CD, Node.js, TypeScript, MongoDB, GitHub Actions Case study: https://www.siddhvasudev.com/projects/ios-release-engineering-signing-testflight-and --- ### Personalised Recommendations With a Closed Feedback Loop Client work · Founding engineer (contract) — engine, integration and audit · in development A recommender for a consumer shopping app, built across the AI service, the backend and the iOS app so that what a shopper does actually changes what they see. #### The detail The engine is stateless and returns product IDs only: a time-decayed taste vector built from likes and passes, up to three taste centres instead of one blurred average, explore slots kept apart so they can be measured, weighted negatives that fade, and diversity limits by brand and category. Cold start uses the text side of the same image model. The backend hydrates results through the existing audience guards and falls back to the previous ranking on any timeout or failure. Before claiming anything I audited the code as it would ship, across all three services, and found 39 issues. The worst: the app re-sorted the server's answer, so a swipe could never visibly change the feed. The fixes closed the loop with a batched product-events endpoint, impressions joined to the response that produced them, and items shown repeatedly without engagement treated as soft negatives. Live testing found what review could not: a release filter that matched nothing, and a cold vector search slower than its budget, now warmed at app launch. It runs on a test environment; no engagement uplift is claimed until it is measured. Python, CLIP embeddings, MongoDB vector search, Node.js, TypeScript, Swift. #### What I did - Built a stateless content-based engine with multi-interest taste, explore slots, decaying negatives and diversity limits - Integrated it through the backend with audience guards and a fallback to the previous ranking - Audited the shipped code across three services (39 findings) and fixed the loop end to end - Added a batched product-events endpoint with impression attribution - Found and fixed two failures only live traffic showed, and stated that no uplift is measured yet #### Stack Python, Recommender Systems, Vector Search, CLIP, MongoDB, Node.js, TypeScript, Swift, Machine Learning Case study: https://www.siddhvasudev.com/projects/personalised-recommendations-with-a-closed-feedback --- ### A Linux Desktop Shell: Dynamic Island, Task View and Live Widgets Developer tools · Creator, sole engineer · personal project A desktop shell for a tiling Wayland desktop, written in a week, that gives Linux the polish of macOS and Windows 11 and a home screen that answers "what needs me right now". #### The detail A dynamic-island menu bar grows from compact to a full hub and carries notifications, the launcher and control-centre panels. A dock minimises with a genie effect, a task view shows live previews of every desktop, and a custom lock screen unlocks the moment the right password is typed while PAM still makes every decision. The desk holds 29 glanceable widgets fed by the machine's real state: agents at work, pull requests, live servers, dictation, the calendar, commits and a daily activity diary, plus insight widgets drawn from the day's agent sessions. Widgets rearrange like phone widgets. The rules I held to: one theme switch drives every surface, input-driven UI changes in the same frame, the shell never moves the pointer or steals focus, and every visual change is checked from real captures. Quickshell (QML), Hyprland (Lua), Luau, GLSL shaders, Node.js, systemd. #### What I did - Wrote the shell on Quickshell and Hyprland: island, launcher, dock, task view, lock screen and start-up curtain - Built a 29-widget desk with a widget library, drag-to-arrange and a Node data feed - Themed every surface from one switch, including Qt and GTK apps - Ran the shell as systemd user services so it restarts on failure - Shipped 173 commits in seven days, each visual change checked from real captures #### Stack Linux, QML, Lua, Luau, GLSL, JavaScript, Node.js, systemd, Hyprland, UI Design Case study: https://www.siddhvasudev.com/projects/a-linux-desktop-shell-dynamic-island --- ### Work Map: A Whole Body of Work on One Zoomable Board AI products · Creator, sole engineer · personal project An infinite canvas that shows a hundred-plus projects, two jobs and a job search on one board, kept current by a daily job. #### The detail Every project, role, idea and technology sits on the canvas as a card with one colour per status; zoomed out, the cards become coloured tiles. A personal layer of notes and stickies syncs across devices with compare-and-swap writes. A daily updater reads every git repository, every GitHub repository and every folder opened in my agent tools, merges duplicates, asks a model to place anything new once, drafts suggested edits from the week's commits for review, and deploys. Client repositories never appear by name, and their commit messages are redacted before the model sees them. It stays smooth at about 1,800 cards: 60 fps while zooming and dragging. The site sits behind a password gate that fails closed, tested against forged cookies and open redirects. Vanilla JavaScript with no build step, Python, Vercel, Claude, Playwright. #### What I did - Built the zoomable board in vanilla JavaScript with no build step - Wrote the daily updater that reads every repository and agent workspace and drafts edits for review - Redacted client names and commit messages before any model call - Kept it at 60 fps with about 1,800 cards - Locked it behind a fail-closed password gate and tested the lock #### Stack JavaScript, HTML, CSS, Python, Vercel, Claude, Playwright, Automation, Data Visualization Case study: https://www.siddhvasudev.com/projects/work-map-a-whole-body-of --- ### Moving a Working Dev Machine to Linux With Every File Verified Developer tools · Sole engineer · complete Moving a laptop from Windows to Linux when its disk holds 533 git repositories and worktrees, irreplaceable datasets and fourteen months of agent history, all recorded with Windows paths. #### The detail A survey sorted every file on three drives into tiers. A single-command backup to an external drive fingerprinted everything with SHA-256, and an install-day runbook had go and no-go gates. Scripts then brought the work back: re-clone what is on GitHub, overlay uncommitted files, restore unpushed commits from bundles, and rewrite every Windows path in the coding agents' saved state, tested first on a full copy of the real data. 135,218 files were verified on restore, with a full insurance mirror kept beside them. Daily jobs moved to systemd timers afterwards. Python, PowerShell, Bash, SHA-256 manifests, git bundles, systemd. #### What I did - Surveyed three drives and wrote a tiered move plan - Built a single-command backup with SHA-256 manifests and an insurance mirror - Wrote restore scripts for clones, uncommitted files and unpushed commits - Re-pathed the coding agents' saved state, tested on a full copy first - Verified 135,218 files on restore and ported daily jobs to systemd timers #### Stack Python, Bash, PowerShell, Linux, Git, Automation, systemd, Data Integrity Case study: https://www.siddhvasudev.com/projects/moving-a-working-dev-machine-to --- ### F2F: A Fitness Coach Agent That Edits Its Own Plan AI products · Creator, sole engineer · personal project A personal fitness companion whose coach agent changes the plan itself when real life gets in the way. #### The detail An intake chat writes a plan from a knowledge base that is the only allowed source of numbers - published dietary and activity guidelines, with evidence labels on traditional practice - then tracks the days and adjusts weekly. The coach has visible, deletable memory and edits the plan directly: "gym four days, evenings" rewrites the week, and "this exercise hurts my knees" swaps in a similar movement and says why. Ask for a one-off workout and the app draws and animates it at once. Every exercise is drawn procedurally on a jointed figure, from a catalogue of 152 movements, and a guided player with set counters and rest timers works offline. Node with TypeScript run directly, SQLite, a PWA in a Capacitor Android shell, and a coach running on Claude or Codex with schema-checked plans. #### What I did - Built the intake, plan, tracking and weekly adjustment loop - Restricted every number to a cited knowledge base - Built a coach agent with visible memory that edits the plan itself - Drew 152 exercises procedurally on a jointed figure with a guided offline player - Verified 85/85 screen checks and 5/5 memory scenarios with the real agent #### Stack TypeScript, Node.js, SQLite, PWA, Capacitor, AI Agent Development, Canvas, Claude Case study: https://www.siddhvasudev.com/projects/f2f-a-fitness-coach-agent-that --- ### Echo: Phone Capture to a True-Scale 3D Model of a Room AI products · Creator, sole engineer · in progress An Android capture app and a Linux pipeline that turn a walk around a room into a measurable 3D model, using a phone with no depth sensor and no cloud service. #### The detail The app locks exposure, white balance and focus for the whole scan, keeps only sharp full-resolution photos, stores each with ARCore's camera pose in metres, and guides the walk with live hints and a coverage map. On the laptop, COLMAP solves the camera positions again from the photos, ARCore's poses scale them to metres, OpenMVS builds a textured mesh for Blender, and Brush trains a Gaussian splat for photo-real viewing. The first real scan was 219 photos of my room. Checked against 18 photos held out of training, the splat's renders scored a median of 29.4 dB PSNR. The accuracy check against tape measurements and merging several rooms are still to come. Kotlin, Camera2, ARCore, COLMAP, OpenMVS, Brush, Docker, Python. #### What I did - Built the Android capture app with locked exposure, white balance and focus, sharp-frame selection and a coverage map - Recorded ARCore camera poses in metres with every photo to give the model true scale - Built the laptop pipeline from photos to a textured mesh and a Gaussian splat, in Docker on the local GPU - Checked the splat against 18 held-out photos: median 29.4 dB PSNR #### Stack Kotlin, Android, ARCore, Computer Vision, 3D Reconstruction, Python, Docker, Gaussian Splatting Case study: https://www.siddhvasudev.com/projects/echo-phone-capture-to-a-true