Omar Cusma Fait

About Me

While pursuing my bachelor’s degree, I developed familiarity with complex problems in math and physics, analytical thinking, and problem-solving skills that have proven valuable in approaching challenges from a logical perspective. I give my best when I’m working on projects that I am passionate about. My experience with chess taught me to stay calm under pressure and to look for a better solution for any given problem.
I am passionate about reinforcement learning, physics, AI, programming, and quantum computing.

Education and Training

[2023] Bachelor’s degree in physics at the Università degli studi di Trieste
[2023] Internship in astrophysical data analysis at the INAF (supervisor: Elena Orlando)
[2015] High School diploma at the "Istituto di Istruzione Superiore Gabriele D'Annunzio"

Language Skills

• 🇮🇹 Italian (Mother tongue)
• 🇸🇮 Slovenian (Mother tongue)
• 🇬🇧 English (Proficient, B2)
• 🇯🇵 Japanese (Elementary)

Digital Skills

• Python (in-depth: Bachelor’s + personal projects)
• Python libraries: NumPy, Matplotlib, PyTorch, Scikit-Learn, Pandas
• Obsidian, Overleaf
• R (Data Science and AI Master's)
• Fortran (Physics Bachelor's)
• Pascal (Secondary school)
• Krita (self-taught image editing)
• Windows, Excel, Word, PowerPoint, Ubuntu

My Projects

Semantic Retrieval (Academic Project, 2024)

With this system, you can find files in a given directory based on the actual meaning of your queries, rather than keywords. It leverages the semantic embeddings generated by powerful LLMs.
Collaborator: Andrea Scomina.

Survey of Image Classifiers (Explainable AI Project, 2024)

Survey of various image classifiers, and the resilience of GradCAM to perturbations. This project investigates the robustness of popular computer vision models trained on the ImageNet dataset—AlexNet, ResNet50, Vision Transformer (ViT), and Swin Transformer—against adversarial attacks and perturbations. It also inspects the reliability of the explanations these models generate under such conditions.
Collaborator: Luis Palacios Flores.

Royal Game of Ur (Reinforcement Learning Course, 2024)

Implemented a superhuman AI for the 4,000-year-old Royal Game of Ur using AlphaZero-style self-play reinforcement learning. The agent learns exclusively through self-play, achieving strategic depth without human data.
Explore the implementation.

Mancala Self-Reinforcement Learning AI (Personal Project, 2022)

I created a self-reinforcement learning AI for the game of Mancala. The agent uses 2-ply exhaustive search and a neural network-based value function to choose the best move. The value function itself is trained from scratch, leveraging self-play. During each generation, a huge batch of games is created, and a new model is trained to predict the future expected reward in each board position. Throughout the training process, deeper search, bigger datasets, and better models are employed to ultimately reach a model with excellent performance.

tisext – Intelligent Time Series Extension (Personal Project, 2025)

A Python library that extends periodic time series by learning underlying patterns, even in noisy or interrupted signals. Unlike naive repetition, it models the true signal structure, supports batch processing, and includes optional plotting for visualization. I tested it on a real-world dataset to forecast product demand.
Install via pip install tisext or view on GitHub.

Mini-Hold'Em (Coursework Project, 2025)

In this project , I tested how evolutionary algorithms perform in mastering hidden information card games of various degrees of complexity. Mini-Hold'em is a simplified version of Poker Texas Hold'Em. The bot is able to achieve human-level play for simpler versions of the game, but it struggles when variability is too high.

Semantic Dynamics – Physics of Language Models (2025)

In this article, I combine Natural Language Processing and Semantic Embeddings to show how a sequence of tokens passed through a semantic embedding can be interpreted as a particle moving through a potential landscape, enabling the evaluation of thermodynamic quantities like temperature, energy, and entropy.
Read Paper

Telegram Bot (Personal Project, 2025)

A multi-functional Telegram bot I built for personal use, featuring utilities like automated Secret Santa assignments, reminders, and custom interactions for my friend group.
GitHub repository.

Wordle Solver through Entropy Minimization (Personal Project, 2025)

A strategic helper for Wordle that selects guesses to maximally reduce uncertainty. Built on my search heuristic framework, it evaluates word candidates by minimizing expected entropy after each guess.
See the solver in action.

Wiki-Tok (Personal Project, 2025)

This group project implements a recommendation system for Wikipedia pages. The user can provide feedback for each page, and then he is fed with the next one, just like with the app TikTok.
Collaborators: Andrea Longo , Stefano Tumino , Alessandro Minutolo .

Search Heuristics (Independent Study, 2024)

In the rich and well-studied field of optimization, one generic use case is to find the best one among a set of random variables, trading off between reward and exploration cost. This project provides a simple implementation and comparison of some of the best-performing heuristics, ready to be applied in real-world scenarios. A very similar approach was implemented for AlphaGo, as it had to select the best move to play based on the results of the Monte Carlo tree search.

Group Theory (Independent Study, 2024)

In this project, I computed the subgroups of the S_4 group. Classifying subgroups involves understanding group theory fundamentals, like subgroup structure, element orders, normality, closure checks, and isomorphism. It also involves recognizing structural equivalences between different-looking subgroups and employing computational tools like SymPy to verify the results.

Quantum Logic Gates (Personal Project, 2021)

In this project, I tested my knowledge of quantum mechanics by creating a module for simulating quantum logic gates. The library enables the user to visualize the process of applying quantum operators and perform simple quantum algorithms.

Quantum Ground State Simulation (Independent Study, 2022)

Using my knowledge of quantum mechanics, I created a repository that allows you to simulate the ground state of a quantum system of interacting particles. Given the potential, the program will find the ground state of the system using the variational method. The program makes the distinction between bosons and fermions. It can also handle potentials with a singularity at the origin.

Ramanujan Machine (Independent Project, 2019)

The Ramanujan Machine is a tool based on continued fractions and gradient descent. It allows finding interesting equations for important mathematical constants. I created a GitHub repository that successfully implemented this tool, and it got featured in a YouTube video.

Peak Detection (Independent Study, 2021)

I created this simple CoLab program to help a friend in automatically localizing and estimating the intensity of peaks in real-life spectrometric data. It applies a kernel to preserve the signal (up to the 3rd Taylor expansion order), and to remove statistical noise. Then, a 2nd order fit is performed around every data-point, to localize the peaks. Finally, clustering is applied on the peaks to remove redundant results.

Steganography (Independent Study, 2021)

I created a repository that allows you to explore how steganography affects the individual bits of an image. The scheme that I implemented overwrites the least influential bits in the image, but only in high-entropy locations, to blend the changes with background noise.

Neural Networks from Scratch (Independent Project, 2019)

This repository, which I created from scratch, can be used to build and train a neural network. This effort tested my knowledge and understanding of the subject, particularly NN architecture and gradient descent.

Wave simulation (Independent Project, 2019)

I created a simple simulation of a 2D wave using the Runge-Kutta method. The library allows for different boundary conditions, initial conditions, and obstacles that the waves interact with. The simulation can be exported to a GIF.

Impulse (High School Project, 2015)

My first attempt at a proper simulation. Built in Pascal, it had a proper user interface where you could manage your simulations. It was capable of simulating gravitational attraction, repulsion, collisions, and each particle could be programmed to follow certain movements, like stochastic walks, chase other cells, and even reprogram them.

Other Projects

Shimanami – Narrative Role-Playing Game (2025)

A personal passion project: a full-fledged narrative tabletop RPG inspired by games like D&D, set in a unique world blending mystery, philosophy, and exploration. The system emphasizes collaborative storytelling and character-driven arcs.
🏝 Explore the world of Shimanami!.

New Beginning – Token-Based Game System (2025)

A lightweight, modular game engine designed for fast-paced, token-driven gameplay. Built as part of the Shimanami universe, it offers a fresh alternative to traditional dice mechanics.
🇬🇧 Regole New Beginning
🇮🇹 New Beginning Rules

Game Design & Development (Personal Hobby)

In my free time, I enjoy designing and developing new card games and board games. This collection showcases various game mechanics, rule systems, and prototypes that I create as a creative outlet combining strategic thinking with game theory.
Explore my games collection!

Hackathons

Generali Challenge 2025

Developed a recommendation system for optional insurance coverages that truly enhance customer protection. Focused on explainability and personalization.
Collaborators: Sepehr, Maddi, Vale, Paolo
Project repo

Hackathon Cybertec 2025

Built a global demand forecasting model using time series regression to predict real-world product demand. Handled multi-horizon, multi-region forecasting under uncertainty.
Collaborators: Uros Zivanovic, Luis Fernando Palacios, Liza
Project repo

Aindo Hackathon 2024

Collaborative hackathon project focused on data-driven innovation.
Collaborators: Uros Zivanovic, Gabriele Pintus, Andrea Ivkovic
Project repo

Google Hash Code 2019

Participated in Google’s worldwide algorithmic programming competition, solving complex optimization problems under time constraints.
Collaborators: Nicolò Rossi, Simone Mezzavilla, Kleit Rrokja
Solutions archive

Other Facts About Me

[2015] Participant of “Scienze under 18”, an in-person activity that aims to bring math to the attention of the public and young students (supervisor: Caterina Vicentini)
[2007-2010] Awarded 22 Chess medals in official tournaments