Empirical ML researcher studying model uncertainty, failure detection, and iterative computation.

I build reproducible PyTorch experiments with controlled baselines, multi-seed evaluation, calibrated metrics, and explicit reporting of negative or inconclusive results.

My background is in theoretical physics, Monte Carlo methods, numerical modelling, and scientific computing. I am currently interested in model evaluation, agent failure modes, uncertainty estimation, and AI safety.

Tests whether hidden-state trajectory instability detects language-model errors beyond predictive entropy. Two particle-seed evaluations produced AUROC differences of +0.00017 and -0.00031, providing no current evidence of a robust advantage.

A holomorphic neural contour-flow prototype with analytical Jacobians, exact-reference validation, and collapse-aware importance-sampling metrics. A preregistered 3-architecture × 3-seed experiment produced a uniform negative result caused by held-out importance-weight collapse.

A compact language model using repeated shared latent-state updates. In a three-seed experiment, four shared updates improved validation BPC from 3.008 ± 0.090 to 2.501 ± 0.019; learned parallel branches were slower and performed worse.

• Master's graduate in theoretical physics
• Python, PyTorch, NumPy, SciPy, C and C++
• Monte Carlo methods, numerical optimisation and uncertainty estimation
• Native English and fluent French