Applied AI Engineer shipping a production LLM agent that turns healthcare providers' free-text WhatsApp replies into validated structured records at 95% field-level accuracy, with a deterministic finite-state machine in control and Claude/OpenAI (swappable behind an LlmPort abstraction) used only for language understanding.

Kept hallucinated or malformed data out of production by revalidating every extracted field before persistence and gating safety-critical refusal behavior with a standing eval harness of 120 adversarial scenarios; deterministic disambiguation cut LLM calls on ambiguous replies by 70%.

Took the agent to live traffic through a WhatsApp Business Solution Provider (Zenvia) and pg-boss workflows across a 6-service NestJS/TypeScript architecture on GCP, built a Python RAG pipeline for Spine under ANS/BCB/ANPD controls, and delivered the Next.js front end for both products.