AI / GenAI engineer — RAG, multi-agent systems, and the plumbing that makes them hold up.
Final-year B.Tech Data Science · Navi Mumbai, India

I build retrieval and agent systems end to end — the spec, the architecture and the code come from the same person. Most of what I ship runs offline or at zero external API cost, because the problems I've worked on were air-gapped or personal-data-sensitive.

I also build the evaluation in from the start. Signer-independent cross-validation, Recall@5 tracking, test suites — a number I can't reproduce isn't a number I'll quote.

Currently: final-year B.Tech, open to AI / GenAI engineering roles and freelance work.

The system I run my other projects through. Work is dispatched to rooms, each with a manager seat and implementation workers, and the model is chosen per role, not per person: DeepSeek v4 Pro builds and reviews, deepseek-reasoner manages, v4 Flash does the cheap graph and summarisation work. Room task envelopes are bounded, so a worker's full transcript never enters a manager's context — which is what stops a long build from silently becoming a context-window problem.

Underneath it is a state layer that had to be made honest the hard way. A session row read RUNNING for 39 days because nobody ever wrote a new value; liveness is now derived at read time, never stored. Token spend is counted from real transcript usage, deduped by message id — one API message spans several log lines, and counting each one triples the bill.

55 test modules. Native components run subprocess-isolated, so a crash costs one utterance and never the control plane.

Repo is private — it ingests my personal notes. Happy to walk through the architecture.

Grounded Q&A over technical manuals. An open-source rebuild of the offline agentic-RAG architecture I shipped to a client at TCS, re-targeted to public Wärtsilä 32 marine-engine manuals so the whole thing is publicly runnable. Rewrite → route → retrieve → grade over an embedded Qdrant index with cross-encoder reranking and an optional Neo4j cascade-fault graph. Every answer is citation-grounded — never invented. Five-step quickstart; no Docker, no GPU.

A personal knowledge engine that turns the firehose you already read into decisions you act on. Watches repos, papers, news and launches, ranks them against your topics, and pushes a digest to your phone. Share a reel and it transcribes locally, researches it, and returns a do/skip verdict. The point isn't more to read — it's less to read and something to do.

Turns a business's public digital footprint into a defensible reason to make a specific phone call. The rule the whole system enforces in code is no receipt, no claim: every fact carries a provenance tier and a source URL, there is no path that writes an attribute without one, and a fact nobody published is null rather than false. Deterministic scrapers with robots enforcement feed a provenance graph; a rubric turns signals into a fit score with a written reason. 145 tests, and zero LLM calls in the collection layer — a rerun produces the same records, so a bad parse is a bug you fix rather than a temperature you tune.

Private — it holds real businesses' contact data.

The same idea pointed at my own pipeline: sweep listings across boards and company portals, rank them against a profile, draft the application, and track what actually happened to each one. Exposes an MCP server so the drafting step can query the pipeline instead of being handed a pasted job description.

Private — it holds my applications and other people's contact details.

Live sign-language → English captions on Meet / Zoom / Teams, as a Chrome extension that runs entirely in the browser — webcam video never leaves the machine. Evaluated under 5-fold signer-independent cross-validation (no signer appears in both train and test): a 128-sign shippable vocabulary at 0.64 mean per-sign accuracy, 115 of 128 signs clearing a 0.50 floor.

nucdesal

Could India's planned 100 GW nuclear fleet desalinate seawater with its reject heat? A cited physics model, 415 tests, every reference number independently re-derived — five didn't hold up.

real-estate-splat

3D walkthroughs of flats from an ordinary phone — IMU pose + monocular metric depth + fusion, then a three.js viewer. 137 tests.

Paper World

15 paper-trading agents on live Binance data, with a planted control group so the swarm's "learning" is falsifiable.

Get Your Hands Right

Inspect a 3D model with your bare hands via webcam — MediaPipe landmarks drive rotate/scale/place, plus a real-time filter stack.

BUDDY

Ducted-fan aircraft engineering rig — parametric OpenSCAD CAD plus Python propulsion, structural and mass solvers, behind a build-readiness gate.

facelessVideos

Script → voice → B-roll → subtitles → rendered 9:16 short, end to end.

Repair Chatbot

Offline RAG over 18,995 repair manuals — the earlier, simpler ancestor of the Agentic RAG Engine.

Portfolio

Cinematic scroll-driven personal site — live.

ugh-boardroom

UI prototype: a multi-agent boardroom where four personas deliberate a decision.

Tata Consultancy Services — AI Systems & Automation · Jan – Apr 2026 Shipped an air-gapped agentic RAG assistant to 350+ fab engineers at an offshore semiconductor client, replacing 10,000+ pages of OEM manuals and cutting troubleshooting from ~40 min to <10 min per incident, at zero external LLM API cost. Recall@5: 0.61 → 0.89.

Reliance Jio — Data Science Intern, Network Analytics · Jun – Jul 2025 Churn models across 20+ cohorts drawn from a 440M-record base (0.838 ROC-AUC, XGBoost); distributed EDA over 200+ variables in Spark.

portfolio · linkedin · [email protected]

_{last boot 14 Sep 2026, 02:49 ist}