Valdera
Shipped Valdera’s first production AI system: an LLM supplier-discovery pipeline on Cloud Run replacing a manual workflow that took hours per request. Co-owned the V2 redesign on LangGraph.
Cut inference cost ~10x — $80 to $8 per request — by re-architecting on an agentic framework and trading retrieval volume for precision.
Owned the LLM platform layer. Migrated off Vertex AI to OpenAI behind a schema-validated wrapper: structured outputs, tool use, prompt caching, retry and backoff.
Built the eval harness and observability that made quality measurable, then cut model output reject rate from ~70% to ~30%.
Automated ~600 inbound emails a week of manual ops triage with an LLM classification and workflow system.
Replaced a third-party search vendor with a custom retrieval stack after coverage proved inadequate for Chinese chemical manufacturers.