Past projects

This master's thesis presents a novel mathematical framework for aggregating individual moral value systems into a consensus system for AI alignment, introducing two principled aggregation methods—Value-Oriented Aggregation (VOA) and Action-Oriented Aggregation (AOA)—and analyzing their theoretical properties and real-world applications.

Delphi is a set of tools for standardized and reproducible training of small language models, enabling seamless integration with Hugging Face for dataset handling, tokenization, and model training.

Implementation of computer vision, reinforcement learning and navigation algortihms for an inter-university competition of flying two autonomous drones for disaster relief on indoor environments.

This project explores the use of Convolutional Autoencoders, GANs, Diffusion Models, and VAEs for watermark removal in images, comparing their effectiveness through quantitative metrics and visual evaluation.

This project explores reinforcement learning techniques, implementing Q-learning for grid-based navigation and deep Q-networks (DQN, DDQN, and PER) for Atari games, analyzing their performance and strategies.