I build AI products
that work in the field.

A chat-driven, audit-rigorous SDK for authoring Avni implementation bundles: a deterministic generator wrapped in a Claude-Agent loop where every turn is a git commit and a validator runs after each one.

16 skill modules behind in-process MCP tool-gates the agent MUST call; validator-truth injection so it can't hallucinate error codes; defense-in-depth safety (prompt-injection output filter, path-jail, auto-revert of any commit the agent makes outside the server); and a real-LLM eval harness that scores agent quality release-over-release.

• ›Built and open-sourced the SDK (164 commits, actively maintained); the skills library it generates was forked into the company org and is actively developed by various internal teams.
• ›463/463 unit tests + a real-LLM eval harness (quality, release-over-release).
• ›Hard cost ceiling: $5/session, $1/turn, thrash-detector — a typical build costs ~$0.13.
• ›Deterministic generation in ~1.7s; the generator 'brain' deployed live on Railway.

A Claude-Code workspace that brings support knowledge — distilled from 90k+ Freshdesk tickets and years of production data-fixes — to the point of work. A support engineer can triage a ticket, investigate it against read-only production data, author a reviewed data-fix, and get reporting help, all in one place.

Four commands — /triage, /investigate, /report, /datafix — over a strictly read-only / draft-only architecture. Freshdesk is read live and replies are drafted for a human to send; queries hit a read replica (RLS-scoped hot standby); /datafix writes a .sql file a human reviews and applies. The production write database is never connected.

• ›Built solo and adopted by the support org — 90k+ Freshdesk tickets distilled into an assistant at the point of work.
• ›−60% ticket resolution time, +40% team efficiency — on real production tickets.
• ›Read-only / draft-only by construction — the prod-write DB is never wired in.
• ›Every fix is a human-reviewed .sql file; the AI drafts, a person applies.

The AI system that turns an NGO's programme spec into a deployable, field-ready mobile app — forms, validation rules, and logic generated by RAG + LangGraph, then human-reviewed.

RAG over organizational docs + LangGraph orchestration generate the config; shipped behind CI/CD with New Relic observability. Recognised as a Digital Public Good; selected for Tech4Dev's AI Cohort 2.

• ›Ships to Avni — 70+ organisations, 11M+ beneficiaries tracked, 100M+ field visits.
• ›Ekam Maternal Health: implementation cut from 2–3 weeks to ~1 day.
• ›30–50% effort reduction per implementation.
• ›One pilot: 8 visit types, 4 programmes, 32 forms, ~300 rules — generated automatically.

A production RAG assistant I built as a freelancer for a fintech client — retrieval over domain docs driving an LLM through tool-calling, streamed live, with a compliance guardrail on every response — then handed to their team to run and extend.

pgvector retrieval → LLM with 37 tool-calling functions → real-time streaming over Server-Sent Events. An output-safety layer scans every response against domain rules before a user sees it. Migrated to serverless edge, cutting hosting cost ~12×.

• ›Built solo, 0→1, and delivered to the client's team for maintenance and further development.
• ›37 tool-calling functions orchestrated behind one assistant.
• ›Compliance guardrail on 100% of responses.
• ›~12× infra cost reduction via serverless migration.

A production investment platform for an AMFI-registered mutual-fund distributor — investor onboarding, a compliance-first disclosure layer, and encrypted nominee records — built solo and running live.

Next.js 15 + PostgreSQL on Railway with auto-deploy from main. Two separate auth paths — email/password for admin, JWT with password-reset or OTP for investors. Nominee PII is encrypted at rest with a re-encryption path for key rotation. The disclosure layer renders ARN/EUIN, the DPDPA-2023 data-protection notice, and the grievance-officer channel on every page. Inbound leads dispatch to Discord on a 5-minute cron.

• ›Live in production, serving a real AMFI-registered distributor.
• ›Compliance-first by construction: no NAV forecasts, no guaranteed-return claims, ARN/EUIN and grievance channel on every page.
• ›Nominee PII encrypted at rest, with a re-encryption path for key rotation.
• ›Built and shipped solo, 0→1 — architecture through production operations.

A trilingual e-catalogue for a Guwahati stone retailer — 1,611 products, 1,608 of them imaged — where every navigation link is derived from the catalogue rather than hand-listed, so the site structurally cannot ship a dead link.

Next.js + Payload CMS on Railway, fronted by a Cloudflare Worker. Navigation columns are generated from live category counts and the image manifest — a column with nothing in it is omitted whole, so a dead link can't be authored. Trilingual (English, Hindi, Assamese) with locale-aware routing. WCAG 2.5.5 tap-target compliance is enforced by Playwright specs in CI, not by review.

• ›1,611 products live, 1,608 with imagery, browsable in three languages.
• ›Structurally dead-link-free: every nav and footer column is derived from real catalogue state.
• ›WCAG 2.5.5 (44px) tap targets enforced by automated specs, not spot-checks.
• ›Built and shipped solo, 0→1, including the production domain and CDN architecture.

A hybrid deterministic + AI tool that turns a 309-item School Eye Health Rapid Assessment (SEHRA) PDF into a full analysis report — quantitative scoring by code, qualitative classification and narrative summaries by Claude, with a human-in-the-loop review before publishing.

Deterministic PDF parsing (pdfplumber) + a codebook scoring engine handle the quantitative side; Claude does only the fuzzy work — classifying free-text remarks into 11 themes and writing summaries + action points. Streamlit app with a human-in-the-loop review step, Dockerised for any cloud. Used for programmes in Liberia, Laos, and India.

• ›Cut a ~2-month manual analysis to 2–4 hours (per report, per analyst).
• ›Deterministic parsing + scoring; Claude only for classification & summaries — accurate and cheap.
• ›Scales from 'can't keep up with 6/year' to 100+ analyses/year.
• ›Runs for < $250/year; human approves every report before it publishes.

A responsive video-streaming interface I built in React — full playback, a related-videos rail, and a real-time live chat — engineered so the UI stays fast and cheap to run under continuous load.

Aggressive client-side caching, debounced input, and a disciplined state-management layer cut API traffic by 92%; the live-chat panel renders new messages at sub-35ms latency through memoised, render-optimised components. React + TypeScript end to end, tuned as a frontend performance exercise.

• ›−92% API calls via caching, debouncing, and disciplined state management.
• ›Real-time live chat rendering at sub-35ms latency under continuous message flow.
• ›Playback, related videos, and live chat in one responsive layout across breakpoints.
• ›Built solo, 0→1, as a frontend performance and real-time UI showcase.

OpenCam (ProCam Studio) — an offline-first camera and screen-recording studio that runs entirely in the browser, with no server and no upload. Record from the camera, the screen, or both at once, pick your devices and resolution, and keep every clip on-device.

Built directly on Web Platform APIs — MediaDevices for camera and mic selection, the Screen Capture API for screen and combined recording, and IndexedDB for a fully offline media gallery. Capture resolution scales from 480p to 4K; the whole thing is a responsive, cross-browser React app with no backend at all.

• ›Camera-only, screen, and combined recording — with camera + mic device selection.
• ›480p–4K capture, a local media gallery, and full offline persistence via IndexedDB.
• ›Zero server: all capture and storage happen on-device, nothing is uploaded.
• ›Responsive and cross-browser, built solo on Web Platform APIs alone.