Scientific machine-learning researcher working on structure-preserving models, learned dynamics, and uncertainty in complex physical systems.

My background is in theoretical physics, including quantum field theory, particle physics, nonlinear dynamics, solitons, Monte Carlo methods, and scientific computing.

I develop reproducible PyTorch experiments with controlled baselines, multi-seed evaluation, calibrated metrics, and explicit reporting of negative or inconclusive results. My current interests include:

• learned contour deformation and sign problems in lattice field theory;
• self-supervised world models for scientific systems;
• adaptive mathematical inductive biases;
• hierarchical and iterative latent representations; and
• uncertainty and failure detection in learned scientific models.

Tests whether perturbation-derived hidden-state instability identifies token errors beyond predictive entropy. Two particle-seed evaluations gave AUROC differences of +0.00017 and -0.00031, providing no current evidence of a robust advantage.

The project includes deterministic data splits, unperturbed controls, calibration metrics, checkpoint resumption, and locked evaluation procedures.

A PyTorch research prototype for learning holomorphic contour deformations of oscillatory complex integrals. The implementation uses analytical complex Jacobians, exact low-dimensional references, and collapse-aware importance-sampling diagnostics.

A preregistered three-architecture, three-seed experiment produced a uniform negative result: all variants eventually underwent held-out importance-weight collapse. The project therefore identifies a failure of the shared phase-focused objective rather than claiming a successful solution to the sign problem.

A controlled study of repeated shared computation in a compact language model. Four applications of a shared latent operator improved validation BPC from 3.008 ± 0.090 to 2.501 ± 0.019 in a three-seed Tiny Shakespeare experiment. Learned parallel branches were slower and performed worse, providing a useful negative result for that mechanism.

This project motivates my interest in iterative and hierarchical latent dynamics, while remaining explicit that the current benchmark tests next-character prediction rather than reasoning.

• Two completed second-year Master's programmes in fundamental and theoretical physics
• Quantum field theory, particle physics, nonlinear dynamics, and solitons
• Symbolic FFTLog infrastructure for redshift-space galaxy skew spectra during a research internship at LAPTh
• C++ Monte Carlo implementation for pure SU(2) lattice gauge theory
• Numerical work on Casimir energies and stochastic annihilation systems

I am interested in combining these lines of work through adaptive structure-preserving world models. One possible direction is a JEPA-like scientific model whose predictor contains a holomorphic component and a general residual, with an explicit penalty measuring when the model departs from complex analytic structure.