Primitives over frameworks. Systems over hype.
I'm Francis. I build AI platforms and the engineering cultures that run them. This is where I think out loud about both.
writing — recent traces
delegation-lab: minimum sufficient oversight, in code
A research paper's closed-form governance math, implemented as a pure functional core with FastAPI and Typer adapters — and validated empirically by a mean-field simulator.
I did not just build a reviewer — I built the thing that measures it
A harness-first AI code reviewer: structured, scored output on one side, and an eval harness with planted issues and false-positive traps that proves the reviewer actually works.
work — measured output
| 40B+ tokens/month on an internal LLM platform, scaled from 3B — gateway, orchestration, evals | 1,500+ engineers served by the platform and an AI-native SDLC transformation | 800+ servers migrated to AWS EKS — founded the SRE function from zero |
| 22+ engineers led across four squads: AI foundation, core AI, enabling, engineering practices | ~50% measured productivity uplift from enterprise Copilot rollout, via DORA and SPACE | 1st author, IEEE paper in computer vision · B.Sc. UFMG · exchanges at CMU and UTEP |
operating principles
Own the primitives
Frameworks are borrowed opinions. When the system is critical, the abstractions must be yours.
Measure or it didn't happen
Evals, DORA, SPACE, cost per token. Every claim on this site has a number behind it.
Trust is the real platform
Business, compliance, and engineers adopt what they trust. Architecture earns trust; slides don't.
Boring infrastructure, bold bets
SRE discipline underneath, aggressive AI experimentation on top. In that order.
about
I spent nearly eight years at one of Brazil's largest mobility companies, going from systems administrator to AI & Engineering Practices Manager. Along the way I founded an SRE function, moved a data center to the cloud, and built the internal LLM platform that now runs a meaningful part of the company's engineering work.
I write here about the intersection I care most about: AI platforms and engineering culture. The technology is the easy half.
Based in Belo Horizonte. Open to hard problems anywhere.