Things that run outside a paper: models, a product, demos and tools. The research behind most of them is on the research page.
Open European language models, Villanova.AI, November 2026
The second generation of Villanova, released in November 2026: a dense 2.4B and a 15.5B mixture of experts with 3.9B active parameters, a 131k-token context, and base, instruction-tuned and vision-language versions, for English, Italian, German, Spanish and French. Open weights and documented open data. All four cite Multi-SFT, our paper on permissively licensed multilingual instruction data.
Product I run
Paste a job posting and Tildea writes a resume for it from your own profile, then checks every claim against what you actually said you have done and removes anything you did not. It scores the result the way an applicant tracking system reads it, and handles cover letters and interview prep for the same job.
Open European language models, Villanova.AI, 2025 to 2026
The first Villanova models: a 2B model released as a preview in December 2025 and as Villanova-2B-2603 in March 2026, with base, instruction-tuned and vision-language versions for the same five languages. I built Villanova-SFT-2603, its instruction-tuning data: 1,711,114 conversations, curated, deduplicated and language-checked, released under Apache 2.0.
Deployed industrial system
Models that predict the carbon monoxide and nitrogen oxide output of gas turbines, served as an API handling over a million predictions a day, which cut tuning time by a fifth. Two papers came out of it, one on the models and one on adaptive forecasting.
Research system, four languages
Turns a passage of course material into a crossword: pick the answers, write a clue for each with a fine-tuned model, then build the grid. It runs in Italian, Arabic, Turkish and English, and the clue models and data are public. The playable crossword on this site was made the same way.
Demo on Hugging Face Spaces
Give it a piece of educational text and it splits it into subtopics, writes questions for each, simulates students with different profiles answering them, grades the answers, and prescribes homework with a short study summary per student. It runs on your own OpenAI key.
Python package
Turns a time series into a graph, as a natural visibility, horizontal visibility or quantile graph, so that graph neural networks can classify signals such as ECGs. It came out of my master's thesis.
Annotation tools
The interfaces native speakers used to rate generated material: Arabic crossword clues for the Arabic crossword paper, and Persian multiple-choice questions for PersianMCQ-Instruct. Released so the human evaluations can be repeated.
Open code and models
Every paper with code has one repository and one model collection, cross-linked, with the datasets and adapters that produced the numbers in the tables. Most are LoRA adapters, so the results can be reproduced on a single GPU.