I build full-stack and AI systems used by 12,000+ people.

Learning to code usually means three tabs: a chat window, somewhere to run code, and an algorithm visualizer. Nothing connects the teaching conversation to what actually executes on screen.

• Simulations from BST up to system design stay in sync with the code tracer's line highlighting. SM-2 spaced repetition schedules reviews from how you actually performed, not just what you clicked. The complexity profiler curve-fits measured runtimes against the target O-class instead of trusting the model's claim. Docs moved out of a 2.88 MB client bundle into server-rendered JSON entries, backed by six RLS-enabled Supabase migrations. About 600 tests in 81 suites are green.

Citizens had no easy channel to report civic issues or understand local ward-level governance and representative data.

• Citizens can file an issue against their actual ward and see who represents them. Resolution status is public, so follow-up no longer depends on knowing the right office. Admins triage through a dedicated moderation workflow. Under the hood it's a Vite + React SPA on Supabase (Postgres with row-level-security policies and Edge Functions), deployed on Vercel, with ward-boundary data loaded from a GeoJSON dataset and rendered with MapLibre GL.

Students had no single place to go. Everything lived across dozens of WhatsApp groups and manual processes.

• Adopted as the primary campus platform within one semester. Redis caching cut database load by about 70% (measured with pg_stat_statements), keeping responses under 200ms at peak. One codebase builds for web, native Android, and an Electron desktop app, with a GitHub Actions pipeline running Jest unit tests and Playwright E2E checks on every push. All of it is run and operated by one person alongside a full course load.

India's gig workers are excluded from formal credit because delivery/rideshare platforms issue no portable, verifiable income record.

• You upload earnings screenshots or PDFs, tesseract.js OCR and per-platform parsers extract the numbers, and an optional Groq LLM pass fills gaps before a review step where you correct anything it got wrong. The result lands on an aggregated dashboard and exports as a SHA-256-hashed, shareable PDF whose public verify link detects edits after issuance. Auth is Google OAuth plus email OTP, data sits in Supabase Postgres behind row-level security, and the node:test suite passes on real OCR fixtures. Account Aggregator Phase-2 is stubbed.

Land-record errors block loans, sales, and inheritance for millions of Indians. Finding out what's wrong means navigating opaque government portals first.

• Automates the real Karnataka RTC portal end to end, walking its District, Taluk, Hobli, Village, and Survey cascade, verified against the production site. Parses Gujarat AnyROR PDFs privacy-safely, deleting uploaded files right after parsing, with a manual-entry fallback and printable application packets. It grew out of an earlier blockchain concept after field research showed the real gap is workflow, not data integrity.

In cyclones and floods, communication channels collapse and control rooms overload. Victims need one-tap help without creating accounts, and coordinators need triage they can actually trust.

• Data lives in a ten-table Supabase schema with row-level security and realtime sync. Around 40 admin routes sit behind JWT auth, with Twilio and WhatsApp broadcast as fallback channels. An offline-first PWA queues reports in an IndexedDB outbox (bridged by BroadcastChannel) whenever connectivity drops. Flood polygons come from a Sentinel-1 SAR pipeline with Otsu thresholding and feed straight into evacuation routing. On top of that: pHash anti-fraud for damage photos, a psychological-first-aid chatbot module, and native BLE mesh clients for dead zones. Live v3 on Vercel, fully open source.

Eligible families miss out on benefits they're entitled to because criteria are fragmented across opaque portals and nobody can tell them what applies.

• A 408-line deterministic eligibility engine keeps entitlement decisions auditable, so qualifications are never invented by a model. The LLM layer only adds deadlines, documents, and news. There's an SSE-streamed staggered agent pipeline underneath, district data covering 740+ Indian districts, plus a what-if simulator, document vault, application tracker, and guided quiz.