Production-grade agentic AI. I build AI agent systems that can be trusted near real money — LangGraph agents behind deterministic guardrails, orchestrated sagas with idempotency keys and transactional outboxes, crypto-shredded PII, end-to-end identity with OIDC/Keycloak, durable execution, and chaos tests that prove all of it.

Not "I made a chatbot." The difference is everything that stops an LLM's surprise from becoming an incident report.

🌐 pablogodoy.com · ✍️ Blog: the production loan-agent series · 📫 godoypablom@gmail.com

A full conversational loan-acquisition agent — LangGraph + MCP on top of a durable, event-driven banking core (TypeScript/Node + Python), running entirely offline with one docker compose up:

The article series walks it layer by layer:

• Sagas under an AI agent — retries that can't double-write, an outbox that can't lose an event
• Guardrails — five deterministic rings the LLM cannot route around
• PII crypto-shredding — "right to erasure" as a one-key operation; the AI layer never sees PII
• Identity — real OIDC/Keycloak from chat UI to tool call
• Agentic RAG — a two-memory, human-reviewed loop that teaches while it sells
• Testing — a deterministic fake LLM brain + Toxiproxy chaos, driven by ~77 e2e scenarios

LangGraph · AWS Bedrock / AgentCore · MCP · TypeScript / Node · Python · PostgreSQL · event-driven architecture · DBOS durable execution · XState · Keycloak / OIDC · Docker · evals & chaos testing