At Vaiu AI, I engineered a RAG pipeline over 6,637 chunks from 40 documents, with a BM25 fallback, and measured 117.7 ms p50 uncached latency across 200 real queries.
I also built a LangGraph supervisor that routes turns among builder, knowledge-base RAG, and Composio tool agents using deterministic intent rules. It reached 49/50 correct on a routing test suite without an LLM call.
For Chat-Builder, I owned the workflow layer, designing Pydantic schemas and a validator that turns LLM-generated JSON into executable React Flow graphs. I also fixed a Pipecat transition race and added a LiveKit TTS cache that cut Google Cloud TTS cost 95% on bilingual phrases.
Outside Vaiu AI, I built a yard-management system for Bee Pee Coatings and contributed GDPR and agent-exit controls to Ever Gauzy. My projects include a published Pipecat session-continuity library and a GitHub App that checks OpenAPI schema changes.
