Software engineering lead at BNP Paribas CIB, with 10+ years building production systems and applied R&D software. My independent ML work focuses on learned representations, model behavior and controlled experiments.
I'm interested in how models represent a changing world, retain information and learn from experience. My recent work is in interpretability and geometric deep learning; I'm exploring world models, memory and reinforcement learning as the next research direction.
I'm interested in Research Engineer and Research Scientist opportunities, as well as Machine Learning Engineer and Data Scientist roles involving substantial modeling and experimentation.
SAE context and intervention prediction. Tested whether activation context predicts the effect of editing an SAE feature in Gemma 3. The observational study covered 12,000 documents, the intervention pilot tested three features. Contextual differences were present, but the fitted predictors did not beat a training-mean baseline on pooled check data. Study and results.
SAE Feature Atlas. A Python library and CLI for collecting SAE activations, inspecting token contexts, and comparing feature statistics, coactivation and decoder geometry. Built to support the study above, with reusable collection and analysis components.
Geometric deep learning on QM9. Compared topology features with geometric and constant controls for molecular HOMO-LUMO gap prediction using an EGNN. In a matched four-condition, three-seed experiment, the topology-conditioned variant performed worse than the controls. The next requirement is a stronger reproduced baseline. Results.
My professional work includes trading analytics, pre-trade risk, data pipelines and distributed systems. Earlier industrial computer-vision work led to a co-authored conference paper in 2018. I hold an MSc in Software Engineering with honors and I am a member of MIPT Deep Learning School.
For ML experiments I use Python, PyTorch, NumPy, pandas, scikit-learn, Transformers and SAE-Lens. My engineering background includes Java, C++, C#, KDB+ and SQL.