Projects
Piggy
Piggy is a web application designed, developed, and maintained independently, intended for self-hosted installation.
The application was born from a personal need for managing personal finances (wallet management), combined with the lack of applications that are sufficiently easy and intuitive, even for those unfamiliar with the topic.
Piggy allows manual, facilitated, and customized tracking of transactions across any number of accounts, with an interface designed to be as pleasant and modern as possible. The application supports various categories of expenses and income, as well as payees and transfers, along with integration with icon and company logo providers.
The project was developed using Laravel and Vue.js, with the support of Tailwind CSS.
Piggy (Android)
In addition to the web application, a mobile client was also developed for Android operating systems using Android Studio.
The application heavily leverages Google's Material Design and closely follows the design of the web application. The companion client is designed for the immediate logging of transactions on-the-go.
Character animation
I also enjoy creating simple 3D computer graphics, mainly using Blender, constantly striving to improve for my personal satisfaction.
For the Fundamentals of Computer Graphics course at the University of Brescia, I had the opportunity to develop a complete 3D character model (Johan Liebert, from the popular manga and anime Monster), modeled from references using box modeling and high-poly sculpting techniques, completed with a skeleton through a process of rigging, skinning, and finally animated. The entire project was rendered using the Cycles rendering engine.
The final animation is available here.
PoIAna
PoIAna is a project organized into two parts (Reinforcement Learning agents and a Godot / C# GUI) that allows anyone to play the popular Italian card game Briscola against autonomous agents trained with the latest AI techniques.
Using Deep Reinforcement Learning, agents for playing Briscola were trained in the Gymnasium environment, refining intelligent and efficient reward functions through heuristics and experiments.
Finally, the models were exported from the PyTorch backend in ONNX format, then imported into .NET / C# to integrate them into a GUI built with Godot, creating a simple autonomous card game.
Want to give it a try? Go to my Itch webpage.
Deep normal map generation
As part of the Image Data Analysis course, I independently developed a simple project for generating/predicting a normal map from a 2D texture.
During the project, various neural network architectures were developed for an image-to-image translation pipeline, ranging from simpler models (Autoencoder and U-Net) to more complex and specialized ones (Pix2Pix and CycleGAN), comparing their performance. The training, construction, and evaluation of the models were carried out entirely independently.
The entire project was implemented in Python using the Keras framework.
Website
I have also developed fully functional websites, including the one you are currently visiting and the one for Raffa's Tobacconist No. 2 (website in italian; it was entirely built with Laravel, Vue.js, and Tailwind CSS).
CBT (C# Behavior Trees)
CBT is a DSL (a domain-specific language) designed to facilitate the creation of behavior trees, a common pattern in game AI.
The transpiler compiles CBT source code into a set of C# classes ready to be used with game engines such as Unity or Godot.
As a demo, the behavior tree was implemented to autonomously control Pac-Man, also reconstructing the original game environment in Godot, watch it here.
RNNMIDI
The RNNMIDI project involved designing a neural network architecture based on simple RNN LSTM models, using the PyTorch framework, to generate MIDI music from an initial stub.
Some examples of generated files can be found in the repository with the training code.
LLM4Logs
LLM4Logs is a modern and experimental project that involved using local and open-source LLMs for analyzing system logs related to cybersecurity.
Specifically, models such as Llama, Mistral, Gemma, Deepseek Coder, and Codellama were compared, enhanced through both fine-tuning on the Colab platform and local zero-shot and few-shot learning.