I’m Felipe Ribeiro: founding engineer at a stealth AI startup, founder of MedAlvo, and an aerospace engineering student at ITA. I like systems where a wrong answer costs real money, so I build them tofail loud, not silent.
Products with paying users, open-source tools people install, and the experiments in between.
An adaptive prep platform for Brazilian medical residency exams, run as a self-driving company: AI agents in cloud VMs scrape, OCR, deduplicate and QA the question bank and handle daily operations.
Turns Markdown + LaTeX notes into realistic handwritten pages for iPad, Goodnotes, Notability and reMarkable. A writer model varies every glyph while a legibility clamp keeps it readable.
Remote roles across San Francisco, São Paulo, Belgium and New York, mostly putting LLMs into production.
Building Milo, a peer-to-peer personal AI for teams. Each employee’s agent runs in its own Linux microVM.
Production LLM agents and full-stack systems for a Brazilian fintech, where a wrong answer costs real money.
Credit models for emerging-market B2B buyers that drove automated decisions and seller pricing.
An LLM chatbot over Salesforce CRM data that replaced analyst-mediated forecasts for US & Canada sales.
Years of physics, math and robotics olympiads, 2017 to 2020. That’s where I learned to like hard problems.
Aerospace Engineering, Data Science minor
Most of my code is written with a fleet of agents: Claude Code and Codex running in parallel git worktrees, on a remote box, inside tmux. I spend my time on specs, review and evals.
I ship eval-first. If a change touches a model, it doesn’t merge until it passes suites built from real production cases. Anything that has to be exact, like prices, money and credentials, stays in deterministic code and out of the model’s hands.
~ tmux attach -t fleet agent-1 feat/streaming ✓ evals passing agent-2 fix/retry-backoff ✓ evals passing agent-3 chore/qa-batch running…
Illustrative. The agent box sits behind a Cloudflare tunnel on this same domain.