During my summer internship at the HYQU laboratory at ETH Zürich, I developed a stabilization system for a Cryostat Fabry-Perot interferometer experiment. This involved deriving differential equations for amplitude attenuation and designing an optimal system using pendulums, springs, and eddy current dampers.
Used transfer learning and lottery ticket hypothesis pruning to create an optimal heavy neutral lepton classification neural network. For this project I was part of the Excellence Research Internship Program at EPFL.
Explored Reinforcement Learning techniques to optimize virtual car racing performance in the Gymnasium CarRacing environment. Employed Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO) models, enhanced with reward augmentations, demonstrating the impact of architectural choices and training methodologies on achieving high-speed, stable racing behavior.
DC-powered lamp that uses computer vision to detect notebooks, stop, and turn the lights on. The lamp uses a Raspberry Pi and a TensorFlow Lite model that we trained. Each part was specially modeled and 3D printed or laser cut.
This paper presents a new hardware aware neural architecture search pipeline for Bragg Peak Spectroscopy experiments
This paper presents a new hardware-aware neural architecture search pipeline for low latency deployment
Benchmark and dataset of over 680,000 synthesized neural networks to evaluate GNN- and transformer-based surrogate models predicting FPGA resource usage and latency
Integrated framework that automates the discovery and optimization of neural networks focusing on FPGA deployment without requiring time-intensive synthesis