Independent researcher and engineer: local AI, measured, and science with verified tools. Chercheur et ingénieur indépendant : l'IA locale, mesurée ; la science, avec des outils vérifiés.
- chimere — Rust-native MoE inference runtime with custom CUDA kernels for Blackwell GPUs (speculative decoding, multi-tier memory), targeting a 35B MoE model on a single RTX 5060 Ti 16 GB.
- strixhalo-verified — Qwen3.8-Flash-Next 125B on one AMD Strix Halo: 54 patches and agent-loop speed-ups, checked against full-precision reference logits.
- ramp-quant — RL-guided adaptive mixed-precision quantization for GGUF models.
- chimere-odo, chimere-studio — orchestration and a local-first workspace around the runtime.
- yang-mills-lean-statement — a candidate Lean 4 statement of the Clay "Yang–Mills existence and mass gap" problem, with compiled guards against fake models; v1.1 adds the q-form of the gap and its closure under limits. DOI 10.5281/zenodo.23644079. A statement, not a proof.
- su2-heatbath-gpu — SU(2) lattice gauge Monte Carlo on AMD and NVIDIA GPUs, byte-identical across vendors. DOI 10.5281/zenodo.23644086.
- crossed-cosmos — an earlier exploratory preprint, kept as a historical snapshot with dated integrity corrections (see its README).
AI agents (Claude, GPT, Kimi, DeepSeek) write most of the code and text under my direction. Results are kept only when they pass mechanical checks: preregistered criteria, mutation tests, byte-for-byte reproducibility, the Lean kernel. Limits and negative results are reported alongside the positive ones.