Projects
Systems I actually want to talk about.
Things I've built that survived past the initial prototype phase. The embarrassing half-finished weekend projects are safely hidden.
Internal platform · 2024 - Present
Agentic AI Platform
A core agentic AI platform. A production RAG API and agents API that enterprise products are built on top of — not a wrapper around someone else's SDK.
- Multi-level tenant hierarchy (Organization → Environment → Tenant → Knowledge Bases → Documents) with hard isolation at every layer.
- Full document ingestion pipeline: PDFs, audio, and images routed through OCR (Mistral), transcription (Deepgram), structured parsing, chunking, and embedding into Weaviate — with async SQS processing so the API stays fast.
- Four production agents: Answer, Chart, SQL, and Voice. Multi-model from day one — Claude, GPT-4, Azure OpenAI, Gemini, and OpenRouter are all swappable at runtime.
- Distributed tracing with OpenTelemetry and async job status via webhooks. The boring stuff that makes it actually observable.
Enterprise — confidential · 2024 - 2025
Enterprise Workspace Intelligence Platform
A workplace intelligence platform that unifies Microsoft 365, Google Workspace, Gmail, Notion, and Slack into a single searchable brain — with voice recording, transcription, and an AI assistant that actually knows what happened in last Tuesday's meeting.
- Delta sync architecture across 6 enterprise data sources so the system stays current without re-ingesting everything from scratch on every run.
- PWA offline recording: MediaRecorder with 2-second chunked captures, IndexedDB storage, S3 multipart upload, and a Service Worker for background sync. Works when the network doesn't.
- LLM evaluation pipeline across 7 dimensions — task extraction, knowledge retrieval, user memory, decision tracking, response quality, tool selection, and voice interaction — because shipping AI without evals is just hoping for the best.
- Twilio phone number provisioning, unified semantic search across all sources, PostgreSQL + Weaviate for hybrid retrieval.
Enterprise — confidential · 2024 - 2025
Enterprise Java Platform
A large-scale enterprise Java platform spanning microservices, payments, scheduling, and a full React frontend — delivered across two phases for an enterprise client.
- Two-phase delivery: core platform first, then Payments, Square, Calendly, and self-serve onboarding.
- Java 21 Spring Boot microservices across 4 modules (common, context, integration, realtime) with Keycloak RBAC, Stripe, AWS CloudWatch, and Sentry.
- React 19 frontend with TanStack Query, Jotai, FullCalendar, Framer Motion, and Nango for 3rd-party API sync.
Enterprise — confidential · 2024 - 2025
Healthcare Intelligence Platform
End-to-end healthcare platform spanning a Windows background service, E2EE data layer, ingestion pipeline, FastAPI backend, and web frontend. Five layers, one engineer.
- Built a Windows background application for continuous passive data collection — the kind of requirement that makes most web engineers quietly close the laptop.
- End-to-end encryption across the full data pipeline. In a domain where a breach isn't just embarrassing.
- Sole engineer across all five layers: desktop service, ingestion pipeline, backend API, data layer, and frontend. No one else to blame when something breaks at 2 AM.
Core system · 2024 - Present
Voice AI Platform
A low-latency voice platform for inbound and outbound AI calling that actually sounds like a conversation.
- Real-time WebRTC streaming with live transcription, agent personas, and multi-channel routing across voice, WhatsApp, Slack, and embedded widgets.
- Built as a SaaS foundation — not a demo, not a proof of concept. Something you can bill customers for.
Diet Paglu
An AI nutrition tracker for Indian food that doesn't require you to manually input every grain of rice.
- Zero-friction logging through photos and voice. Because nobody wants to fill out a spreadsheet before eating dinner.
- Uses conversational AI to guess macros and track progress so you can just eat.