AI engineer in Paris, working on ML and systems engineering: LLM agents, their evaluation, and the infrastructure that runs them in production. My foundation is in data science and engineering. Site: sbrouck.org.
Research projects
- rfm-cooccurrence-recency-probe: controlled reproduction of a KumoRFM-2 claim. Tests whether a hosted relational foundation model captures row-level co-occurrence and dependence on the most recent row.
- weight-space: gradient-free weight edits that remove thinking loops from small local models, with MLX.
Open-source contributions
- TransformerLens #1987 (merged): residual-stream decomposition APIs for SSM and hybrid architectures.
- Hugging Face Datasets #8575 (merged): checkpoint accounting across repeated resumes of async
map. - n8n #43328, #43261: agent runtime and model picker fixes.
- Temporal Rust SDK #1772: workflow-task chunking for updates on uncached workers.