Projects

Built, shipped, in use.

Production products and personal explorations. Screenshots and demos are blurred where needed for compliance.

2026·Personal project · solo build, end to endFlagship

PayOps AI

Multi-agent payment reconciliation. When a company's gateway, orders, ledger, webhooks and settlement disagree, a LangGraph system of parallel specialists (Jev + Gemini) investigates, cites evidence and proposes a fix. Deterministic policy, four-eyes approvals, exactly-once execution and an independent validator decide what actually happens.

  • ▹LangGraph 19-node state machine: Jev plans once, 3 specialists fan out in parallel via Send(), findings are grounded against cited evidence, with durable interrupt / resume for approvals
  • ▹Two models, two jobs: Jev for 6 typed decisions with coded fallbacks, Gemini for reasoning. A Jev-first fast path resolves known faults with zero Gemini calls
  • ▹Models propose, code decides: a closed catalog of 9 actions, a P0–P11 policy engine, four-eyes approvals, idempotent execution, and a validator that re-reads every system
  • ▹11/11 golden scenarios passing live (100% root-cause, action and tier match) for $0.004 total; record/replay cassettes make the demo free and deterministic
  • ▹Hexagonal TypeScript monorepo: React 19 workbench, Express 5 + Socket.IO + pg-boss, one Postgres for data, queue and checkpoints. Hash-chained audit, MFA, ~460 tests
LangGraphGeminiJev (TypeSafe)TypeScriptReact 19Node.jsExpress 5PostgreSQLDrizzlepg-bossSocket.IOZod
Narrated 3-minute walkthrough: investigation, evidence, approval, verification
2026·Personal project · in development, week 3 of 8In Development

Hinglish Clinic Voice Agent

A real-time voice receptionist for small Indian clinics. Callers speak English or Hinglish; the agent answers clinic questions, finds free slots, and books, moves or cancels appointments. It verifies the caller before touching their data and never writes anything they haven't clearly confirmed. Every stage of every turn is timed.

  • ▹Streaming LiveKit Agents loop: Deepgram STT → Gemini Flash → Cartesia TTS, Silero VAD + turn detection, barge-in in about 0.35–0.4 s, with Sarvam and Gemini fallbacks
  • ▹Safety in code, not the prompt: one agent with per-mode tool allowlists, caller verification before any personal data, read-backs built from the stored row, idempotent writes and a DB constraint against double booking
  • ▹Sandboxed evals: 13/13 multi-turn tasks, 0 wrong actions, 217 tests; per-stage p50/p95 latency on every turn
  • ▹Doctor-name recognition on phone audio 61% → 99%; slot lookup up to 1.2 s → 200 ms
PythonLiveKit AgentsDeepgramGeminiCartesiaSarvamSilero VADPostgreSQLasyncpgpydanticpytest
HinglishIn development · demo coming soon
2025·Software Development Engineer · Full-Stack

FAXFlo — Diagna AI

Full-stack AI medical-document platform — clinical inbox, eFax triage, voice scheduling, RPA EHR automation, and analytics. Taken from a company pivot to its first paying US clinic.

  • ▹React 19 frontend: clinical inbox, eFax module, scheduling views, analytics dashboard — per-clinic feature flags, RBAC routing, lazy-loaded routes
  • ▹Node.js backend: 40+ REST endpoints, JWT/RBAC auth, Redis BullMQ queues, Swagger docs
  • ▹AWS pipeline: S3 → SQS → Textract → SNS → Bedrock — 1,000+ docs/day, 99%+ uptime, DLQ + Slack alerting
  • ▹Multi-model AI: Claude (Sonnet) via Bedrock + GPT-4o cross-check + Amazon Nova Pro fallback — ~95% accuracy across 40+ medical document categories
  • ▹SMS: Twilio · Telnyx · Vonage orchestrator with provider failover, opt-in/out compliance, automated reminders
  • ▹Voice AI: VAPI + ElevenLabs outbound scheduling — automated patient calls, cutting manual outreach ~70%
  • ▹RPA: Robocorp · FastAPI · Redis RQ for MDLand EHR automation + React Chrome Extension for on-page control
  • ▹Document engine: Zustand + Immer field-level diffing, S3-hosted PDF preview via @react-pdf-viewer, 80% image compression
  • ▹HIPAA-compliant PostHog analytics dashboard — the demo that closed the first paying US clinic
React 19TypeScriptZustandNode.jsExpressPostgreSQLRedisBullMQAWS BedrockClaudeGPT-4oTextractVAPIElevenLabsRobocorpPythonFastAPIPrismaPostHog
FAXFlo product walkthrough — official demo from diagna.ai
2024·Frontend Developer

VoiceGenie

Generative-AI voice sales platform. Designed and built the entire product solo — campaign management, internal analytics, onboarding UX, ElevenLabs voice integration, and CRM connections. Reached $10K MRR in 11 months.

  • ▹Sole frontend engineer — designed and built the entire platform, dashboard, and marketing site from 0
  • ▹Campaign builder with call objective config (lead gen, sales, appointments) and per-contact live tracking
  • ▹Voice configuration panel: ElevenLabs voice, pitch, rate, AI temperature, tone, engagement style per campaign
  • ▹Post-call analytics: recordings, auto transcripts, emotion detection, AI entity extraction (email, phone, address, intent)
  • ▹Internal R&D dashboard: conversion funnels, drop-off by script section, voice model performance comparison
  • ▹CRM integrations: HubSpot · GoHighLevel · Cal.com — entity data logged back, auto calendar booking on success
  • ▹Live @token script composer — personalised per contact at call time, no hardcoded scripts
  • ▹0 → $10K MRR in 11 months · 30–50% faster page loads via code splitting and caching
ReactNext.jsTypeScriptAnt DesignTailwindReduxElevenLabsHubSpotGoHighLevelCal.com
Official VoiceGenie demo (voicegenie.ai) — UI designed & built by me
2026·Personal projectIn Development

MotionStudio

A real video editor that runs in your browser, built on Remotion — canvas editing, a frame-accurate timeline, keyframe animation, 34 text effects and 18 shader backgrounds, beat-synced shots, and two ways to export: free in your browser or full-fidelity on AWS Lambda. Real auth and cross-device project sync. A deep dive into programmatic video and creative tooling.

  • ▹Full editor — canvas editing, frame-accurate timeline, keyframe animation, 34 text effects, 18 shader backgrounds, beat-synced shots
  • ▹Two export paths — free in your browser (WebCodecs) and full-fidelity cloud render via AWS Lambda
  • ▹Production infra — Supabase (auth · quota · cross-device sync), Vercel Functions, AWS Lambda + S3
RemotionReact 19TypeScriptTailwindZustandSupabaseAWS Lambda
The editor — assets, canvas, live properties, and a frame-accurate timeline
The editor — assets, canvas, live properties, and a frame-accurate timeline · click to zoom
2026·Personal project

ClinRAG

Hybrid RAG over FDA drug labels, with the evaluation harness as the real deliverable. Dense and BM25 retrieval fused with reciprocal rank fusion, answers only from retrieved excerpts, and a RAGAS benchmark on hand-verified questions that decided what not to build.

  • ▹Table-aware chunking (Docling), Cohere embeddings on Bedrock, FAISS + BM25 with reciprocal rank fusion, Claude Haiku 4.5 answering only from retrieved excerpts
  • ▹RAGAS on 12 hand-verified questions: hybrid lifted context recall 0.75 → 0.83 but faithfulness fell 0.98 → 0.88, so it is not a clean win
  • ▹Reranking was scoped out on purpose: the numbers didn't show a gap it would close
PythonAWS BedrockClaude HaikuCohere EmbedFAISSBM25RAGASDocling
ClinRAGScreenshots coming soon