ZK. AI, models & on-chain systems.

I wrote a model-editing engine to study how changes to model weights affect behavior. The PRIME pipeline selects edits and compares the resulting models against the original checkpoints.

I track capability scores and output drift, then publish the selected models, settings and evaluation results. The code uses streaming statistics to keep memory use down.

I designed and trained a 58M-parameter French language model from scratch on an RTX 4060 with 8 GB of VRAM. I built the tokenizer and training pipeline, from pretraining to post-training, to explore reasoning at a small scale. The model uses a <think> scratchpad to work through problems, with arithmetic as one evaluation task.

I built a Rust desktop app for running language models locally. It combines llama.cpp inference with an agent workspace, 30+ tools, MCP support and granular permissions. It supports CUDA, Vulkan and Metal, with a French/English interface.

I built an agent harness that keeps API credentials outside the model's environment. Third-party calls go through TAP, while the tool sandbox has no direct network access. External calls are logged, and credentials stay in TAP.

I co-built EventChain with Etherwave Labs for the RealFi hackathon, where our team won first place in the human.tech track ($5,000). The platform addresses event no-shows through USDC deposits: attendees check in by QR code, and a smart contract redistributes absent participants' deposits to those who attend.