About Inference.

I started Inference because the writing I needed didn't exist. The technical articles I respected were either too academic — citation-rich but disconnected from production — or too commercial: tutorials engineered for organic traffic, with no admitted point of view.

What I wanted to read was a practitioner thinking out loud, in public, with their evidence on the table. So I started writing it. Every post is a record of a real decision: a problem I encountered, the options I evaluated, the choice I made, and the reasoning behind it. Where I was wrong, I say so. Where I am still uncertain, I flag it.

There are no listicles. There is no 'top ten'. There is one author. There is a point of view.

I am a software architect and ML practitioner based in São Paulo. I have spent roughly a decade between research labs and production teams — long enough to have shipped systems I'm proud of, and long enough to have shipped systems I now regret.

I hold a Master's in Computer Science from USP, where my thesis examined inference-time optimisation for transformer architectures. Since then I have worked across financial infrastructure, ML platform engineering and distributed-systems consultancy. I have been on call for production AI systems serving tens of millions of requests per day, and I have signed off on architectures I knew would fail in two years. Both experiences inform what I write.

Inference is organised around three topic areas. Every post carries a primary topic tag; the archive is filterable by topic.

Posts are written in the first person, because intellectual ownership matters. When I write 'I argue', I am putting my name behind the argument. When I write 'I was wrong', I am putting my name behind the correction.

The register is precise. Claims are backed by evidence — benchmarks, citations, or clearly labelled personal experience. Speculation is permitted but flagged as such. Hype is not permitted.

Inference publishes in two languages: English at en.inference.dev, Brazilian Portuguese at pt.inference.dev. Each version is written natively from the same source material — neither is a translation of the other. The voice and structure are consistent across both. The examples and references occasionally differ where local context matters.

I read every email. I do not run paid sponsorships, sponsored posts, or affiliate links. If you want to flag an error, send a correction. If you want to suggest a topic, send a question I haven't answered yet.

> beatriz@inference.dev