Research
This is my public research work so far. I am currently working on new papers in multi-agent AI and vision model training with weak supervision. Of course, more work is being developed at XISTRA, and hopefully it will become public soon.
Multi-Agent Reinforcement Learning for Drone-Based Search and Rescue Missions
Master's Thesis - MSc in Industrial Mathematics, in progress
My Master's thesis studies different multi-agent AI architectures based on reinforcement learning, with particular attention to their stability, convergence behaviour, and collaborative dynamics. The work evaluates these architectures across environments of increasing complexity, designed to resemble search and rescue missions where teams of drones must locate one or more targets in novel scenarios.
The goal is to identify an AI architecture that, with the necessary real-world adaptations, could eventually be deployed beyond simplified environments and inputs. A central part of the work is to distinguish genuine collaboration and generalisation from the mere memorisation of patterns, while analysing what each architecture reveals about learning, coordination, and robustness.
The thesis is still in progress. Once it is completed, I will share it here.
Real-Time Aerodynamic Airfoil Optimisation Using Deep Reinforcement Learning with Proximal Policy Optimisation
November 2025 — Aerospace (MDPI), Vol. 12, Issue 11
This work began as my Bachelor's thesis and eventually developed into a published paper. It applies deep reinforcement learning with Proximal Policy Optimisation (PPO) to optimise aerodynamic airfoil profiles in real time within the context of morphing wings. The approach learns to satisfy both aerodynamic objectives and complex geometric constraints while maintaining low computational cost and millisecond-level optimisation speed.
The code is available on GitHub. It was written while I was still learning a great deal about software development, and much of it was implemented manually at a time when AI coding tools were far less capable, so the codebase is certainly improvable.