I'm a machine-learning researcher and postdoc at the Weizmann Institute, with a PhD from the ATLAS experiment at CERN. Two things I built there run in production: GN2, the jet flavor tagger used across ATLAS (4000+ physicists), whose implementation I led, and the GPU graph construction for track reconstruction, which I co-developed and which cut the slowest step from about ten minutes per event to 69 ms. I also led a multimodal foundation model for particle physics, and I'm first author on HGPflow and calorimeter super-resolution.
My current work is on self-supervised learning. In two first-author manuscripts, I use image simulators where every generating factor is known to test what JEPA-style encoders keep, and whether label-free training signals notice when they lose it. Representation learning and world models are what I want to work on next.
I also teach: lectures and tutorials for Weizmann's Practical Deep Learning for Science course, across four editions between 2021 and 2025.
Weizmann Institute of Science. Extending physics-guided ML toward learned world models.
Weizmann School of Science — for outstanding Ph.D. research.
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.
Seven kinds of collider data in one token vocabulary. A task is a choice of which to provide and which to request, so one model covers 1302 of them.
ATLAS's jet flavor tagger for Run 3. Led the implementation and co-designed the auxiliary objectives that make the network learn the decay structure flavor depends on.
Particle reconstruction as an assignment problem. The network predicts what fraction of each cluster's energy came from each particle, so energy is conserved by construction.
Finer calorimeter granularity from software. One conditional flow-matching model splits each cell into f×f smaller ones and removes the electronic noise in the same step.
Contrastive learning for jets where both views come from the simulator: same collision and parton shower, later random stages run twice. The invariances come from the data-generating process.
14 total View all projects
