I build AI agentsthat hold up in the real world.

Selected work

• Pillion

  1st place · Agora Voice AI Hackathon 2026

  Hands-free voice co-pilot for India's delivery riders.

  • 9 LLM tools that act on the phone: GPS, SMS, calls, SOS
  • Response time cut from 24–37 s to ~1 s
  • Offline crash detection, covered by 94 unit tests

  Code ↗Demo ↗

• EasyEV

  1st prize · Agora Ecosphere 2026

  Voice-first AI that helps people choose and buy an EV.

  • Real-time voice consultation with sub-300 ms barge-in
  • Two AI agents debate EVs live, with an AI moderator
  • 3D virtual showroom and live handoff to a human advisor

  Code ↗Demo ↗

• VaaniDesk

  AI voice receptionist that answers business calls and books appointments.

  • Guardrails: prices and slots come only from the database
  • 30-scenario eval harness gates CI, plus 190+ tests
  • Next.js, Fastify, Postgres, Redis/BullMQ monorepo, ~17k lines

  Code ↗

• DebugMind

  Autonomous agent that fixes failing tests, and has to prove each fix first.

  • Harness guardrails: no edits before a verified hypothesis
  • 12/12 bugs fixed with zero regressions (11/12 without guardrails)
  • Sandboxed probes, full-suite re-runs, a trace of every step

  Code ↗

• WatchTower

  Best pitch · Sheryians Coding School Hackathon 2026

  Uptime and incident monitoring with live status.

  • Scheduler with distributed lock, jittered backoff, flap detection
  • Live status over Socket.IO, Prometheus metrics, public status pages
  • 323 tests across 19 suites, coverage-gated CI

  Code ↗Demo ↗

More work

• Bahi

  Voice-first AI operator for small businesses.

  • 13 agent tools across PayPal, Gmail, Slack, Notion and Jira
  • Policy guardrails hold big payments and block refunds; 266 unit tests

  Code ↗

• Aveniq

  Full-stack e-commerce marketplace.

  • Razorpay checkout with signature checks; stock decremented atomically
  • JWT in HTTP-only cookies, Google OAuth, role-based access; 71 integration tests

  Code ↗Demo ↗