At Innowise, I rewrote the core agent for a multi-agent document validation platform using a deterministic engine and LangGraph ReAct agents. After shipping it to production, manual review time fell by ~60%.
I stabilized an unreliable agent with Langfuse tracing, prompt rework, and hardened failure points, reducing failed validations from ~15% to ~3%. I also built a nightly promptfoo evaluation pipeline to catch regressions before release.
For a financial documents RAG assistant, I owned the pipeline from ingestion and chunking through embeddings and retrieval with MongoDB vector search and a Neo4j knowledge graph. I lead AI engineering practices in a Python department, teach an internal AI Engineering course, and mentor engineers.
