I'm a machine-learning researcher with a PhD from the ATLAS experiment at CERN, where several methods I led are in production — including the primary jet flavour tagger for Run 3 (Nature Communications, 2026) and a CUDA graph-construction pipeline delivering orders-of-magnitude speedup over CPU baselines.
My current focus is extending that work toward learned world models for physical and robotic systems. In the last year of my PhD, I worked with Robert Katzschmann's group at ETH Zürich on the design of a flow-matching latent-dynamics pipeline for soft-robotic control; two fellowship proposals came out of that effort (MSCA top-quartile; ETH AI top-10%). The architectural patterns from particle-physics ML map directly to contact dynamics — that bridge is the foundation for what I want to do next.
Weizmann Institute of Science. Extending physics-guided ML toward learned world models.
Defended at the Weizmann Institute of Science. ATLAS Collaboration, CERN.
For contributions to jet flavor tagging with graph neural networks.
Weizmann Institute of Science — for excellence in physics, awarded to international doctoral students.
Began PhD research on machine learning for collider physics.
ATLAS's primary jet flavor tagger for Run 3. Co-designed the auxiliary-task structure that drives most of the gains over single-task baselines.
Reformulating particle reconstruction as an assignment problem. Conservation laws live in the architecture, not the loss.
Super-resolution and denoising on a graph of irregular detector cells — each low-resolution cell expands into f×f children, predicted by one conditional flow-matching model.
A 4M-style any-to-any model trained on the complete causal trace of particle physics — seven modalities, one shared token space.
Re-simulation as the augmentation — share upstream physics, vary downstream stochastic stages. Invariances come from the data-generating process, not invented transforms.
14 total View all projects
