About
I'm Antoine Lépée. I build the systems around AI agents: the memory they keep, the harnesses they run in, and the checks that let people trust what they produce.
I've been doing this kind of work for more than fifteen years, mostly by crossing borders. I started on the front-end, moved to back-end and business domains, touched data and infrastructure, and spent several years as an engineering manager. I joined Evaneos in 2016 and I'm now an AI engineer there. I like the problems that sit between disciplines, and I like to carry them from the first sketch to the thing that runs in production.
AI is not a recent turn for me. I started in 2022 with a conversational interface that helped travellers say what kind of trip they wanted. That thread led to Trip Designer, to an Evaneos app inside ChatGPT, and to what I spend most of my time on today.
What I'm working on
- Tooling for coding agents. A pipeline of thin, independent layers that takes an agent from the intent of a change to the verification of its output, each layer useful on its own.
- Personal OS. A memory for my agents that I own: plain Markdown, versioned, written during the session, caught up in the evening, consolidated once a week. The first posts here are about it.
- Making closed appliances observable. AHome Assistant integration for Hitachi heat pumps, used by other owners, and the instruments around it to understand what the machine actually does.
- An energy model of the house. A model that learns the physics of a home from its sensors, so heating, solar and storage can be planned instead of guessed.
- How Was Your Day. An app that records and transcribes meetings on the device, separates the voices, and writes a summary where every point links back to the moment it comes from.
How I like to work
Small increments that each deliver something. Tools that give teams more autonomy, not more process. Systems you can inspect: if I can't see what a machine or an agent remembers, decides or consumes, I don't trust it yet. I'm fond of the ideas ofsymmathesyand holacracy: learning happens between the parts of a system, and authority works best close to the work.
The drawings on this site borrow the line of 19th century naturalists' plates. Those drawings had one job: show how a living thing is built and how it works.