I build AI systems that run in production — and don't fail silently.
Production RAG, autonomous email and voice agents, multi-agent orchestration. From backend inference to the deploy gate. The kind of latency-sensitive AI that has to be right when a real customer is on the other end.
Autonomous email assistant for automotive dealerships. Reads inbound leads and replies to customers with business-hours logic, follow-up scheduling, and a full audit trail. Multi-tenant data handling with strict per-tenant isolation and sub-500ms hybrid retrieval.
Real-time voice support: several specialized agents under a supervisor, with streaming speech-to-text and text-to-speech over phone lines. Indexed 1M+ records at 98% retrieval accuracy.
Visual AI workflow builder. A drag-and-drop canvas of 13 node types to compose multi-step AI pipelines with no code. Multi-LLM routing, row-level-secure auth, usage-based billing, and a browser capture extension.
Open-source library for querying databases in plain English across MySQL, Postgres, and MongoDB through one interface. Semantic caching cuts inference cost ~40%; a validation pass guards against SQL injection.
Recommendation engine with transformer embeddings — +28% engagement, +15% average order value. Real-time inference API serving 1M+ predictions a day at 45ms P50, with an automated training-to-deploy pipeline on multiple GPUs.
Multilingual event-booking assistant on GPT-4o with live database tools (search, book, cancel), web search, and SMS. Placed top 5 of 800+ projects at the Smart India Hackathon.
I help teams build the eval and harness layer that catches regressions before users do — and I take on a few hands-on builds a quarter. If your evals pass but you're not sure what fails silently, let's talk.