Projects

Multi-byte decoding for tokenizer-free language models

A one-layer drafting head and a gated verification scheme that make byte-level models generate faster, with output identical to greedy decoding.

Hierarchical Reasoning Models with Iterative Revision

Two language models under 600M parameters, pre-trained from scratch, that exploit a recursive architecture

Formulaic Alphas' algorithms

Letting an algorithm write its own trading signals as formulas you can read. First with genetic programming over prices and social-media sentiment, then with a reinforcement-learning agent.

storiafactory

An AI-powered B2B SaaS platform for the publishing industry, co-founded by me, where I lead product engineering and technical architecture.

Our Memories

An iOS app for couples, with twelve dates to live together, letters sealed for the right moment, songs dedicated to each other and Home Screen widgets. Live on iOS.

Audio Transcriber

A small desktop tool that summarizes audio and video on your own machine with Whisper and local models. An open-source release with a proper interface is planned.

TIAGo as a social guide

A multimodal human-robot interaction system that lets a TIAGo robot guide visitors at a technology fair, with socially aware navigation and a dialogue engine with memory.

OCR post-correction with fine-tuned LLMs

Fine-tuning language models of different sizes to turn noisy OCR output into clean Italian text, using synthetic training data.

Blacklight

A first-person horror game built from scratch on three.js, with no game engine and every animation written by hand.

Probe detection

A YOLOv8 nano detector for inspection probes, built as a technical assignment for Flyability.

Blind blur estimation

Telling from a blurred image alone whether it was blurred by motion, defocus or a Gaussian kernel, and estimating how much, as the first step to restoring it.

Classifying cultural items across Wikipedia languages

Telling whether a Wikidata item belongs to every culture, is typical of one, or exists only in one, from how it appears across 73 Wikipedia language editions.