About

I lead AI engineering at Atlax360. The interesting work starts after the prototype: workflow engines that spam users, document classifiers that flag everything as fraud, and retrieval systems that find everything except the answer.

Right now that means three systems in production. BIFROST handles document intelligence. ORVIAN runs B2B workflow automation. Polaris serves internal knowledge search for support, sales, and product teams. Different surfaces, same unglamorous reliability problems.

My bias is toward context engineering, eval loops, and reliability. Most AI systems do not fail because the model is weak. They fail because the context is messy, the orchestration is vague, and nobody defined what good looks like.

I learned to code at nine with a Pascal book and a dictionary missing half its pages. At twelve I copied a hacker website from Geocities so badly I left their HTML comments in the source for six months. When TensorFlow landed, I got pulled into AI. My first chatbot was a disaster: ask about the weather and it tried to sell you insurance.

Before AI I spent years doing full-stack work for banks and retailers. BBVA, Bankinter, PagoFX, El Corte Inglés, Decathlon. Deadlines were real. Compliance was real. Production bugs were expensive.

My day-to-day stack is Python, FastAPI, Next.js, TypeScript, PostgreSQL with pgvector, Supabase, and Redis. OpenAI and Anthropic APIs when they solve the problem. Not as a reflex.

If you are building production AI and stuck in the unglamorous parts, that is usually the conversation worth having.