FOTBCD
An open, geographically diverse building-change dataset with 223,675 before-and-after aerial image pairs and 960,880 building instances across 28 French departments.
AI Engineer
Turin, Italy
I build and evaluate AI systems across retrieval, computer vision, and data-intensive applications.
My work moves between research and engineering: constructing datasets, training models, testing where they fail, and turning the useful parts into working tools.
Featured work
Projects and research
All projects
MBCTD
A dense multi-label model that maps unchanged, demolished, and new buildings from before-and-after aerial imagery, including replacement sites where labels overlap.
FOTBCD-CC
An open dataset and model for describing change in before-and-after aerial imagery, with 321,200 captions for 223,675 temporal pairs.
Search Any Change
A multimodal search system over 420,175 before-and-after aerial patch pairs from Rhône, a department excluded from model training.
AI Legal Research Platform
A full-stack Italian legal-research application that combined source search, cited AI answers, and persistent conversations in one React experience.
EduGen
An AI study platform for cited chat, hybrid lesson search, document analysis, and quiz generation.
Recent writing
Notes from the work
All writing
Inside FOTBCD-CC: analysing 321,200 aerial captions
An analysis of 321,200 model-generated aerial captions: dataset composition, descriptions per location, language patterns, and candidates for visual review.
Distilling Qwen3.8-Max into a smaller aerial change model
The decisions behind making aerial change captioning affordable: teacher supervision, prompt caching, LoRA adaptation, and quality evaluation.
Training a multimodal model to retrieve aerial change
The design decisions behind teaching DINOv3 and Qwen to connect descriptions of change with before-and-after aerial imagery.
Why Binary Change Detection Isn't Enough
A practical look at what disappears when every temporal difference is reduced to changed or unchanged.