kkuetteHi, I'm Tony Denion.
Online, I'm kkuette.
Independent software engineer and self-taught machine-learning researcher, based in France, working remotely.
I tinkered for years, then learned to program for real in C at the École 42 piscine in 2015. Python came around 2017. Lately: language models that keep their own notes.
Say it out loud in French: cacahuète. Yes, peanut. The name stuck.
Now
One person, one GPU, a model that writes its own memory
I run Fractale, a one-person research lab on language models whose long-term memory is written by the model itself. Instead of a longer context window, the model reads a document one page at a time and keeps a handful of notes in its own learned shorthand. Those notes are a few kilobytes you can save, reload tomorrow, or move into another session.
Everything is open under MIT: the weights, the training code, a research log with exact reproduction commands, and a paper on the mechanism. The first checkpoint's whole pretraining run cost about $320 of rented GPU time.
Day to day, the lab runs on a single RTX 3090, a small rig of RTX 3070 Ti cards for ablations, and a handful of home servers for storage, CI and job queues. The scaling campaign beyond that is supported by a EuroHPC JU compute grant (AI and Data-Intensive Applications Access) on MareNostrum 5, hosted by BSC. For a one-person lab, that grant changes what is possible.
This is a human/AI research pair, and I say so everywhere. Research direction, architecture and experimental judgement are mine. Implementation, experiment execution and write-ups are produced in collaboration with Claude (Anthropic). This page included.
Earlier work
Mostly Python, mostly open
- TradzQAIReinforcement-learning environment for trading: backtesting and live sessions, eight agents (DQN, PPO, DDPG, TRPO…), TensorFlow / TensorForce.
- coinbaseadv-pythonUnofficial Python client for the Coinbase Advanced Trade API.
- coinbasewebsocket-scraperMarket-data collector over the Coinbase websocket feed.
- AutoPICSmall JavaScript automation tool.
- École 42 projects in Clibft, printf, get_next_line, fillit, lem-in, filler. Where the discipline comes from.
Path
- École 42, piscine of 2015Tinkering before that; real programming started here, in C. Algorithms and the peer-learning method: no teachers, projects reviewed by other students, and a lot of nights on a segfault.
- Python, around 2017Small tools first, then data and models. Learned by shipping and by reading other people's code.
- Reinforcement learning and trading botsTradzQAI: agents that learn to trade in a simulated market, then a live one. First taste of training loops that never quite converge.
- FractaleLanguage models with a self-written memory. Preregistered experiments, a public research log, and a first open checkpoint in 2026. Phase 2: teaching the model to use its memory deliberately.