Zürich, Switzerland
About
I work on multimodal models for text, image and video. My PhD: getting vision–language models to hold up in a domain as demanding as art — and working out how they get there.
I’m an AI Engineer at Studio Jadu, where I build multimodal text–image–video models for Studio Jadu’s animation pipeline. The point isn’t a machine that produces a series on its own — it’s tooling that extends what a small team can make, with artists and storytellers directing the work.
I’m also a PhD candidate in Computer Science & Mathematics at the University of Bari Aldo Moro, supervised by Prof. Giovanna Castellano and Prof. Gennaro Vessio at CILab. My thesis is submitted and the defence is coming. The work was funded by a fellowship under Italy’s D.M. 118/23 within PNRR Mission 4, on the analysis and valorization of digitized artistic heritage with AI.
What the PhD is about
Two questions, and they turn out to be the same question from opposite ends.
The first: can vision–language models work in a domain that genuinely stretches them? General-purpose models are trained on photographs and web captions. Art is a hard case — a painting carries style, period, iconography and intent at once, almost none of it labelled, and the vocabulary that describes it is specialist. Getting a model to be useful there means adapting it rather than prompting it and hoping.
The second: when such a model does say something, how did it get there? A fluent description is not evidence of understanding. So a good part of the work is interpretability — using token activation maps to trace each generated word back to the pixels it actually draws on, and finding where the grounding holds and where the model is really running on language priors.
In practice that has meant captioning, text-guided multi-mask inpainting, few-shot segmentation with visual prompts, multimodal retrieval, cross-modal generation from artworks to music, and the interpretability work above. The papers are on the research page.
In 2025 I spent six months at the Digital Society Initiative at the University of Zurich with Dr. Eva Cetinić, on that second question. I’m still based in Zürich.
Before that
Both my degrees are from Bari: a BSc in Computer Science in 2021 and an MSc in Computer Science, AI curriculum, in 2023. My BSc thesis was on automatic text complexity assessment with machine learning — work I later extended over four months with the National Research Council of Italy. My MSc thesis was on automatic artwork captioning, which is where the art thread starts.
Elsewhere
I review for NeurIPS and CVPR, publish code for most of what I work on, and sometimes write things down. If you want to talk about multimodal models, generative video, or research that has to survive a product deadline — get in touch.
Timeline
Aug 2026 — Present
AI Engineer
Studio Jadu Zürich, Switzerland — hybrid
Building multimodal text–image–video models for Studio Jadu's animation pipeline: tooling that extends what a small team can produce, with artists and storytellers directing the work at every step.
Sept 2025 — Feb 2026
Visiting researcher — text-to-image generative models
University of Zurich — Digital Society Initiative Zürich, Switzerland
Interpretability of text-to-image generative models, under the supervision of Dr. Eva Cetinić.
- Training data attribution — working out what a generated image owes to the specific examples it was trained on.
- Sparse autoencoder feature analysis on the vision encoders that condition these models, CLIP and DINO.
Oct 2023 — Present
PhD candidate, Computer Science & Mathematics
University of Bari Aldo Moro Bari, Italy
Vision–language models and multimodal AI at CILab, supervised by Prof. Giovanna Castellano and Prof. Gennaro Vessio, funded by Italy's NRRP Mission 4. Two threads run through it: getting these models to hold up in a domain as demanding as art, and working out how they arrive at what they say. Thesis submitted September 2026; defence expected December 2026.
- ArtSeek — a multimodal RAG system pairing late-interaction retrieval (ColQwen2) with agentic reasoning. Its classification head re-uses the retriever's own vision encoder through multi-task learning, reaching state of the art on artwork attribute classification.
- I Dream My Painting (WACV 2025) — the first work on multi-mask inpainting: several regions filled at once, each from its own prompt, in a single generation pass rather than one region at a time.
- Label Anything (ECAI 2025) — few-shot semantic segmentation driven by visual prompts, state of the art on COCO-20ⁱ among multi-class methods.
- WikiFragments — a Wikipedia-scale multimodal dataset, over 5M image–text pairs across 42M examples.
- Large-scale training on the CINECA LEONARDO supercomputer through an EuroHPC ISCRA allocation.
May 2022 — Sept 2022
Software engineer — research collaboration
National Research Council of Italy (CNR) — Institute for Educational Technologies Palermo, Italy — remote
Natural language processing for education, carrying on from the text complexity work I had started in my BSc thesis.
- Built NLP tooling for dyslexia detection in Python and spaCy.
- Contributed to COURAGE, an EU-funded social media literacy platform, on Django and Vue.js.
2021
Student, Samsung Innovation Campus
Samsung Italy & University of Bari Aldo Moro Bari, Italy
Samsung Italy's training programme, run with the University of Bari. Selected among 25 students and placed in the top five on the final project, winning a scholarship.
- Built SmartGym, an Android app that counts exercise repetitions from the camera with an on-device computer vision model.
Education
Sep 2021 – Jul 2023
MSc in Computer Science — Artificial Intelligence110/110 cum laude
University of Bari Aldo Moro Bari, Italy
The AI curriculum, and the two years where the research direction actually formed: the thesis grew straight out of the coursework and became my first paper.
- Machine Learning
- Deep Learning
- Natural Language Processing
- Computer Vision
- Fundamentals of AI
- Information Theory
Leveraging VLP and Transformer Models to Describe Artworks with ChatGPT Data
Nov 2018 – Jul 2021
BSc in Computer Science110/110 cum laude
University of Bari Aldo Moro Bari, Italy
A broad computer science foundation — mathematics, algorithms, software engineering, databases — with the first machine learning courses that pointed where I ended up.
- Mathematics — discrete and analysis
- Algorithms & Data Structures
- Object-Oriented Programming
- Software Engineering
- Machine Learning
- Databases
Creazione di un Tool per la Valutazione della Complessità del Testo