Generative modelling for the physical world.
MSc Data Science at ETH Zurich. I want generative models to design things that work in the physical world. I work on methods that use physical feedback to help diffusion models discover new, valid designs beyond their training data. In robotics, building the humanoid software stack at the ETH Robotics Club convinced me that reliable automation comes from shaping the robot and its environment around the task, not from one general-purpose machine. Before that, I built DeFi protocols for fixed income and margin trading.
Worked on the software stack for the club's Unitree G1 humanoid, from teleoperated data collection to deployment: NVIDIA GR00T for manipulation, with the SONIC neural WBC for locomotion and navigation, on bin-picking tasks developed with industry partner Cartier. This stack is the foundation being ported to XALUN, the club's fully Swiss-built humanoid.
Co-founded a spot and margin trading venue on Sui, built on DeepBook, and grew it from zero to $3.5M in assets under management. Built the trading backend, with real-time order synchronisation and market-data streaming, and the liquidity vaults. Backed by the Sui Foundation.
Core engineering team of an interest-rate derivatives protocol, from pre-launch to over $250M in TVL. Built the custom AMM that carries the protocol’s trading volume, and the limit-order stack on top of it. Worked on MetaVaults, curator-run vaults that hold any on-chain position with built-in accounting and settlement.
Using generative models, guided by a magnetohydrodynamic equilibrium solver, to discover new and physically valid stellarator designs for fusion.
Implemented DSRL on NVIDIA GR00T N1.6 for the ETH Robotics Club's Unitree G1. An off-policy RL actor learns the noise fed to the frozen diffusion policy, steering its behaviour without touching its weights. Trained on parallel Isaac Sim rollouts for pick-and-place. Code
Built at an SF hackathon at AGI House: a quadrotor policy that intercepts evasive drones from noisy sensor observations, trained by cloning a classical proportional-navigation guidance law. Five rounds of DAgger raised its closed-loop hit rate from 22% to 30% across five target behaviours, against 37% for the expert with ground-truth access. Code
Adapted mimic-video, a video-action model built on NVIDIA Cosmos-Predict, to a low-cost SO-101 arm. LoRA-fine-tuned the video backbone and trained flow-matching action decoders from scratch on up to 8 H100s, then deployed closed-loop on the real arm from a single RTX 5090. Code
Fine-tuned NVIDIA GR00T on the Unitree G1 from teleoperation demonstrations. Training setup and evaluation.
Co-led the first undergraduate-run workshop at Applied Machine Learning Days 2024, with 50+ attendees. Built and taught a neural combinatorial optimisation pipeline in which a policy trained with policy gradients learns to solve travelling-salesman instances, with applications to multi-satellite scheduling. Code · Announcement
Coursework: diffusion models, log-Sobolev inequalities and Markov semigroups, robot learning, vision algorithms for mobile robotics, large-scale AI engineering, guarantees in ML.
Coursework: Brownian motion and stochastic calculus, large language models, mathematics of data science.
English, French, Polish, German