Who I am

I am an AI Engineer with a Computer Science degree from UFF, in Rio de Janeiro. I work at Global Hitss on generative AI in a corporate context and, in parallel, I build LLM products on my own, from requirements to production. Two of them are commercial products, sold and running in client operations.

The systems I have built share one problem: a wrong answer looks exactly like a right one.

A clause the model invented looks exactly like the real clauses in the contract. A query built the wrong way returns a plausible number, in the right format, and nothing on the screen says it is wrong. A financial entry corrected in place erases the version the accountant will ask for three months later.

That problem is what shaped how each of them was built. In the system over tender documents, every statement comes with the passage it came from, and a citation that matches no passage never reaches the screen. In the one that answers questions about the database, the query is built from the real structure of the tables, not from what the model assumes is there. In the financial system, correcting a mistake creates a new entry instead of erasing the old one, and the full history stays there for whoever audits it later. In all of them, the hard part was not making the system answer. It was leaving enough of a trail for someone to check the answer.

The chat on this site works the same way. It answers from the content of these pages, shows the passages it used before the answer, flags on screen any citation that does not match them and, when it finds nothing about the subject here, says it does not know instead of filling the gap. You can check where every statement came from, including the ones it makes about me.