Gabriele                   

My latest project focuses on pixel-level classification of Martian surface imagery into five terrain classes. Built with TensorFlow and a U-Net architecture, it reached a mIoU of 0.675 on the full test set and emphasizes spatial reasoning, mask generation, and visually interpretable output.

Classes

5 terrain types

mIoU

0.675

Model

U-Net

Developed for the Artificial Neural Networks and Deep Learning course, this model performs multiclass blood cell recognition from microscope images. The final system reached 92% accuracy on the hidden test set and focused on building a reliable image classification workflow end to end.

Classes

8 cell types

Accuracy

0.92

Framework

TensorFlow

A complete digital version of MyShelfie built as a distributed systems project at Politecnico di Milano. The game supports multiple simultaneous matches, reconnect flows after client crashes, in-game chat, and parallel GUI and terminal interfaces on top of the same networking model.

Mode

GUI + TUI

Network

RMI / TCP

Architecture

Client-server

An algorithms project modeling a highway as service stations with constrained movement between nodes. The solver computes the route with the fewest stops and applies a lexicographic tiebreaker when multiple optimal solutions exist, keeping the output stable and predictable.

Language

C

Focus

Graph search

Tie-break

Lexicographic

This project compares candidate words against a target by combining positional constraints, frequency checks, and persistent state across guesses. It uses hash tables for fast lookups and Quicksort for ordered output, turning a compact systems problem into a robust word-analysis pipeline.

Language

C

Core DS

Hash tables

Output

Sorted matches