In short

I am a Staff Software Engineer at Jusbrasil, Brazil's largest legal-tech platform, where I build and operate LLM agents that answer legal questions for hundreds of thousands of people every month. I hold a PhD in Computer Science (social network analysis) and have spent over a decade at the intersection of software, data, and law.

I have also taken two products from zero to revenue: a background-check product with tens of thousands of subscribers today, and a WhatsApp legal-intake channel with over 100,000 monthly users. The agent platform I work on exists because of those products, not the other way around.

What I have been doing lately

· Taking a WhatsApp legal-intake agent from experiment to a channel-agnostic chat platform (WhatsApp, web, and app), with a gateway, session cache, and autoscaling on Kubernetes.
· Building the evaluation loop around it: an evaluation service in Go with LLM-as-judge metrics exported to Prometheus, a CLI that runs synthetic users against the agent, and daily eval runs that gate model changes.
· Making agents observable: tokens, cost, latency per tool, and conversation depth per turn, with Grafana dashboards and alerts that fire on quality and engagement, not only on 5xx.
· Working with coding agents as an engineering practice: spec-driven changes, repository skills, and agent instructions that let a small team ship roughly a thousand commits a year with review discipline intact.

My PhD research was about finding experts and trustworthy answerers in question-and-answer communities: who answers well, whom to trust, which discussions deserve attention. The evaluation work I do today is the same problem with the answerer replaced by an LLM. Ten years of measuring the quality of people who answer questions, now applied to agents.

See what I am working on now or the longer background and publications.