GUILLERMO — AI ENGINEER / BUILDS WITH CLAUDE · OPEN TO WORK
Guillermo Hoyo Bravo. GenAI engineer.
RAG and multi-agent systems in production at Gestamp and IDEA, plus a payments-SaaS feature at Zelebrix shipping soon — all built with Claude.
welcome — today’s route is drawnback again — the route’s changed
The model proposes;
the evals dispose.
Selected work
Shipped, in production, and honest about what’s next.
RAG chatbot for a multi-tenant payments SaaS, built and validated, scheduled to deploy in a few months — hybrid retrieval, PII scrubbed before the model.
AI Engineer — sole AI engineer at IDEA; I designed, built and lead the production RAG platform powering two live assistants (HR and electrical-engineering-rules).
Production multi-agent system on Semantic Kernel — Text-to-SQL, document and SharePoint tools, permission-aware.
My lab: this site, built with Claude, plus the experiments that feed my method.
A method, not magic.
Spec before prompt
I write the brief before Claude writes a line, because otherwise it builds the wrong thing precisely and I review the wrong thing carefully.
PREVENTS: confident generation against an unstated spec.
Small verified steps
Claude generates one step; I run it and verify it before the next ask, because otherwise errors compound silently and the bug is three commits deep.
PREVENTS: compounding errors hidden by a big diff.
Evals decide, not vibes
I write the eval set and acceptance criteria; Claude proposes; the tests dispose — because otherwise “looks right” ships and breaks in front of a user.
PREVENTS: shipping on a demo that never met a real case.
SHIPPED
The hardest problem on the Zelebrix build was never the model — it was the privacy boundary. PII is scrubbed before anything reaches the LLM and tenant data is queried locally, so the external model never sees it. The boundary holds by construction, not by a prompt.
About
I’m Guillermo, a GenAI engineer based in Madrid. I hold an MSc in Data Science from the Autonomous University of Madrid, and I spend my days turning automation needs into agent and retrieval systems that survive contact with production — not just a working demo. At Gestamp I extended a multi-agent system and built a Text-to-SQL agent, evaluated against a controlled golden set before every deploy [PLACEHOLDER — Guillermo to confirm: accuracy figure]. At IDEA and Zelebrix I designed RAG platforms and led their builds, reviewed commit by commit. Outside the terminal I’m learning Vietnamese, slowly and badly, which keeps me honest about being a beginner.
NOW (JUN 2026): as IDEA’s sole AI Engineer, leading the production RAG platform; shipping agent work at Gestamp; and hardening the Zelebrix chatbot’s retrieval ahead of its deploy in a few months.
Contact