Portfolio

Projects from my thesis, internship, hackathons and courses, from LLM systems to robotics and high-performance computing.

Trajectories of four underwater agents searching for target zonesMSc Thesis
MSc Thesis · Nov 2025 – Sep 2026

Multi-Agent RL for Underwater Search and Navigation

Teams of autonomous underwater vehicles learn to navigate and find target zones in a physically realistic ocean simulation, as part of the IREG coral-reef project (Sorbonne University Abu Dhabi). Compared PPO, IPPO and MAPPO: communication between agents raised success rate by up to 24% on the hardest targets. Also extended the open-source SwarmSwIM simulator.

  • Python
  • PPO / IPPO / MAPPO
  • Julia
  • Oceananigans
  • SwarmSwIM
LangGraph diagram of the multi-agent RAG workflowInternship
AI Engineer Intern, AI Lab · Jun – Oct 2025

Agentic RAG for Cardiology Guidelines

Multi-agent assistant that answers medical questions on ESC cardiology guidelines. Built the ETL that turns guideline PDFs into a vector index and was the main developer of the chat interface, with authentication and persistent chat history.

  • LangChain
  • LangGraph
  • Qdrant
  • Ollama
  • Chainlit
  • PostgreSQL
  • Docker
Skeleton-based pose tracking of a pedestrian2nd place
Spring Hackathon 2026, Trieste · Mar 2026

The Eye: AI Parking Surveillance

Parking analytics from aerial drone video: occupancy over 364 spaces, dwell times and a pedestrian–vehicle stress index, plus privacy-preserving anomaly detection that only sees skeleton data. Everything comes together in a Streamlit dashboard with a virtual twin of the car park.

  • YOLOv11
  • ByteTrack
  • PyTorch
  • MPED-RNN
  • Streamlit
Advanced HPC course · Jun – Jul 2026

Parallel ICP Point-Cloud Registration

3D point-cloud registration written from scratch in C, with kd-tree search and SIMD, OpenMP and CUDA backends, profiled with Nsight and FlameGraph. Brought the alignment of a 500k-point cloud from 308 s to 3.1 s on CPU, about 100× faster.

  • C
  • CUDA
  • OpenMP
  • SIMD
  • SLURM
Deep Learning course, team of 3 · Jun – Jul 2025

Fish Detection, Tracking and Classification

Pipeline that detects, tracks and classifies fish species in underwater video (F4K, DeepFish), with a real-time mode and a higher-accuracy offline mode.

  • Python
  • YOLOv8
  • ByteTrack
  • PyTorch
HPC course · Sep – Oct 2025

Distributed Heat-Equation Solver

Hybrid MPI + OpenMP stencil solver with non-blocking halo exchange and NUMA-aware memory. Strong and weak scaling study on up to 16 nodes (1,792 cores) of the CINECA Leonardo supercomputer.

  • C
  • MPI
  • OpenMP
  • SLURM
  • Python
NLP course, University of Alicante, team of 4 · Feb 2026

Document Extraction and Summarisation with Local LLMs

Extracts structured JSON from Spanish public-grant PDFs and summarises them with small local LLMs. Summaries are evaluated with BERTScore and an NLI-based hallucination check.

  • Python
  • Hugging Face
  • Phi-3.5
  • Qwen2.5
  • SmolLM2
  • BERTScore
How Gaussian Splatting clones and splits Gaussians during optimisation
University of Alicante, team of 3 · Mar – Apr 2026

3D Reconstruction with Gaussian Splatting

Review of multi-view 3D reconstruction, from Structure-from-Motion and NeRF to 3D Gaussian Splatting. Reproduced the official implementation on Mip-NeRF 360 with a free T4 GPU, matching the paper's quality (PSNR 32.4 dB).

  • 3D Gaussian Splatting
  • COLMAP
  • PyTorch
  • Colab
RL course, team of 3 · Feb – Sep 2025

Chess Reinforcement Learning

Top contributor. A bitboard chess engine in C++ exposed to Python, used to solve endgames with value iteration, Monte Carlo Tree Search and other RL algorithms.

  • C++
  • pybind11
  • Python
  • MCTS
  • Google Test
Optimisation for AI course · Jan – Feb 2026

Co-evolution of Robot Controllers

Two robots fight in a custom 2D energy game while their neural controllers evolve against each other with NEAT (population 512), with matches evaluated in parallel.

  • Python
  • NEAT
  • Neuroevolution
  • multiprocessing
Heatmap of activation patching results across GPT-2 layers
NLP course, University of Alicante · Dec 2025 – Jan 2026

Mechanistic Interpretability of GPT-2

Activation patching experiments on GPT-2 to find which layers and attention heads carry the information behind a prediction.

  • Python
  • PyTorch
  • minGPT
  • Hugging Face
CEVAE model architecture
Probabilistic ML course, team of 3 · May – Jul 2025

Causal Effect Inference with CEVAE

Estimating treatment effects with hidden confounders, from latent variable models trained with stochastic variational inference to the deep CEVAE architecture.

  • Python
  • Pyro
  • PyTorch
  • Variational inference
Advanced Programming course, team of 3 · Nov 2024 – Jan 2025

Scientific Computing Toolbox

C++ library with Python bindings for statistics, interpolation and ODE solvers, installable with pip. I wrote the interpolation module and the CMake build and packaging.

  • C++17
  • pybind11
  • CMake
  • Eigen
  • GSL
In progress
Personal project · 2026

Drone Navigation with ROS 2

ROS 2 drone navigation stack targeting a Jetson Orin Nano: mock drone, controller and mission-manager nodes, with PX4 SITL and a Docker development environment.

  • ROS 2
  • C++
  • PX4
  • Docker
  • Jetson
Cloud Computing course · Mar 2025

VM vs Docker Cluster Benchmark

Built the same 3-node cluster with VirtualBox VMs and with Docker containers, then compared them with HPCC, iPerf3, stress-ng, sysbench and IOZone.

  • Docker
  • VirtualBox
  • Linux
  • Networking
Shared with classmates · 2024 – 2026

SDIC-DSAI Course Notes

Open LaTeX notes for the Data Science and AI master's courses, written together with classmates.

  • LaTeX
  • Open source