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

// why this project?

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.

// what does it do?

An agentic coding tutor on a Gemini → Groq → OpenRouter → offline cascade. Every answer ships with a runnable Monaco editor — Pyodide in a Web Worker locally, or a Judge0 / Wandbox / Piston fallback chain server-side. 17 steppers stay synced to the tracer, tied together by a 59-topic curriculum DAG.

// results

• ▸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, and the Supabase Postgres schema sits under row-level security across 29 migration files, with 78 test files.

// why this project?

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

// what does it do?

College platform at SOA/ITER serving 12K+ students. One hub for events, attendance, timetable, notes, PYQs, and AI grade analytics — replacing scattered WhatsApp groups. Open source, with web, Android, and Electron clients.

// results

• ▸Adopted as the primary campus platform within one semester. A Redis cache layer (with an in-memory fallback) sits in front of Firestore, MySQL and Postgres data stores, and a Playwright importer pulls a student's official portal data after the student solves the CAPTCHA themselves
• ▸credentials are never stored. One codebase ships as a PWA, a native Android app and an Electron desktop app, with 26 Jest test files running in GitHub Actions CI. All of it is run and operated by one person alongside a full course load.

// why this project?

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.

// what does it do?

SIH disaster-response platform built by a team of four: zero-auth SOS, explainable RED/YELLOW/GREEN triage, OSRM evacuation routing, a command-center dashboard, a volunteer portal and an offline-first PWA in English, Hindi, Bengali and Odia. An offline BLE mesh app adapted from the open-source BitChat covers dead zones.

// results

• ▸A React 19 and TypeScript client talks to an Express and Prisma API with 22 services, 17 data models, a WebSocket hub, OTP and role-based auth, rate limiting and 8 backend test files. Triage scoring is rule-based and explainable, and evacuation routes come from OSRM road geometry with multi-point probing so two distinct routes are always offered. An offline-first PWA queues reports in an outbox when connectivity drops. The Python AI engine (damage grading, flood mapping) is a simulation prototype, not a deployed service. Live v3 on Vercel, open source.

// why this project?

Existing Python AI frameworks give up either performance or memory safety. In Rust you don't have to pick one.

// what does it do?

A 9-crate Rust workspace for LLM applications: radix-tree HTTP routing, five LLM providers (OpenAI, Anthropic, Gemini, Ollama, Groq) with automatic failover, a from-scratch HNSW vector index with write-ahead-log persistence, four memory backends, an agent runtime, semantic caching, JWT / API-key auth and PyO3 Python bindings.

// results

• ▸About 13.6K lines of Rust across nine crates (core, llm, memory, agents, vector, prompt, cache, auth, py). The vector crate implements HNSW itself (no external index library) with metadata filters and WAL + snapshot recovery
• ▸memory can be in-memory, SQLite, Redis or Postgres. Criterion benchmarks cover vector operations, and GitHub Actions runs CI plus a cargo-audit security workflow. PyO3 bindings let existing Python teams adopt it incrementally instead of rewriting everything at once.

// why this project?

Posting the same content to several networks by hand is slow, and failed or half-published posts are easy to miss.

// what does it do?

Social media scheduler built during the JJ Software internship: compose once and publish to Facebook Pages, Instagram, X, LinkedIn and Google Business, with queues, best-time slots, evergreen recycling and client approval links.

// results

• ▸Each network is published independently with per-target locking, exponential-backoff retries (30 s, 2 min, 10 min) and a reconnect-required state when a token is rejected. Provider tokens are encrypted at rest behind OAuth 2.0 PKCE, every action lands in a tamper-evident audit log, and a client approval gate rejects edits made after review. Large videos (up to 100 MB) upload straight to S3 with signed tickets. Next.js 14 and Prisma on Neon Postgres, with 27 automated test files.

// why this project?

General-purpose models happily invent case law, and they rarely push back on a weak premise. In legal research one fabricated citation can cost a career.

// what does it do?

Legal research assistant on a 3-agent pipeline: Researcher retrieves, Drafter writes with citations, independent Verifier checks every claim. Weak evidence → abstains. Streams live to UI.

// results

• ▸A deterministic citation guard enforces that every claim maps to a retrieved source
• ▸a score-floor abstention gate refuses low-evidence answers
• ▸a bounded repair loop fixes draft issues before anything reaches the user. Ships offline out-of-the-box (mock model + local embeddings + a bundled US Constitutional corpus), with optional live connectors to Federal Register, CourtListener, and GovInfo. pytest suite plus live-source E2E tests on GitHub Actions CI.