Research

I currently work on large language model inference and post-training as an AI Researcher at a stealth company.

I completed my PhD in Computer Science at the Technion in 2026, supervised by Prof. Alex Bronstein and Prof. Reuven Cohen. My PhD research spanned very different areas: exercise physiology from wearables and video, differentiable physics solvers for wildfire spread, encrypted network traffic, and diagnostics for large language models. Across them I built task-specific deep learning systems for complex temporal and spatiotemporal data, usually noisy, partially missing, and collected outside controlled lab conditions, and I brought domain structure (physiology, geometry/physics, or network behavior) into modern models so they are not only accurate, but also interpretable and reliable in practice. From May to October 2026 I also worked as an independent contractor with Apple’s Video Engineering group, modeling physiological state from 3D human pose in video.

In exercise physiology, I developed methods that estimate VO₂ and related metabolic dynamics, traditionally measured using expensive lab equipment (e.g., Cosmed K5), from lower-friction signals such as smartwatch data and video. A recurring approach in this work was to treat physiology as a dynamical state-estimation problem, using tools like neural ODEs and neural kalman filtering to model how latent physiological states evolve over time and how they manifest in observable signals. Alongside this, I worked with large-scale wearable datasets to support long-horizon understanding of behavior and recovery, including activity recognition under weak supervision and physiologically informed strategies for handling missing data. I also collaborated on applied medical machine learning problems, such as readmission risk and analyses connecting CGM with accelerometer-derived activity patterns.

I also worked on differentiable solvers and inverse problems, where you want to learn from data while preserving meaningful structure. One example is my work on wildfire propagation modeling via differentiable eikonal-equation solvers: the eikonal equation captures wavefront-style propagation, and by making classical fast methods (e.g., fast sweeping) differentiable, typically via implicit differentiation, we can learn propagation-related parameters from observations while keeping the solver grounded in the underlying geometry. This direction appealed to me because it blends learning with structure and interpretability, rather than replacing the model with a black box.

In communication networks, I used real traffic and telemetry to build models that help understand and optimize performance in modern systems. This included analyzing encrypted QUIC dynamics to infer aspects of server–client interactions, and developing learning-based approaches for teleoperation settings (e.g., teleoperated driving), where predicting and adapting to network behavior is critical for reliability and latency.

Toward the end of my PhD, I turned these tools toward large language models, building calibration-free diagnostics that audit what a model actually does rather than what its headline scores suggest. One direction measures the effective depth of the residual stream: how many genuinely distinct computational stages a model performs, as opposed to how many layers it nominally has. Another, the shuffle gap, audits personalization, testing whether a model said to be personalized truly conditions on a specific user’s history or has quietly collapsed onto population-level adaptation. What tied this to the rest of my work was the same lens of structure and effective sample size, applied along the layer and user-support axes instead of time.

Across projects, I like going end-to-end: understanding the problem, building the data pipeline, designing the right model, and evaluating it carefully. I also occasionally provide professional guidance, whether it’s on modeling choices, evaluation design, or turning a research idea into a working system. (See the Publications section for paper links and details.)

Outside of work

I’m married to Shir Cohen, who also holds a PhD in Computer Science. I’m a proud father to Romy and Ilay, and we share our home with Zeus, our loyal German Shepherd. When I’m not spending time with my family, I’m usually running, coaching amateur long-distance runners, or reading about the sport. I’m a proud member of the ASICS Front Runner team and hold a personal best of 2:43 from the Valencia Marathon.