By day I'm a senior AI engineer at ESAP, building real-time voice agents. The rest happens in public: fine-tuning and inference tools on GitHub, the odd Pallas kernel when nothing else will run a model, Kaggle competitions, and long-form writing on JAX, MLX, and where intelligence actually comes from.
The Unsloth experience, for Apple Silicon. Fine-tune LLMs and vision-language models locally on your Mac with LoRA / QLoRA support.
Big open models on Kaggle's free TPU v5e-8, behind an endpoint Claude Code or Codex can talk to. Nothing existing could run GLM-5.3-Flash on a TPU, so it runs on a JAX engine I wrote for it.
A two-part field guide to JEPA — Yann LeCun's bet on learning world models by prediction. First the intuition from first principles (latent-space prediction, collapse, I-JEPA, V-JEPA 2, LeJEPA), then a hands-on half: train and fine-tune these models on your Mac with mlx-tune.
A long-form personal field guide — the JAX paradigm, the hardware substrate underneath (roofline, TPUs, GPUs), the modern ecosystem (Flax NNX, Grain, Orbax), sharded computation, the LLM stack, and writing Pallas kernels. A living synthesis I keep updated as the ecosystem moves.
