AI Engineer · Data & ML
I build systems that find what's similar — and measure whether they're right.
Computer science by training, applied AI by trade. The thread running through it is data that has to be matched, retrieved or predicted, and the question of how you know the answer is right.
May 2021
Max Mara Fashion Group · Reggio Emilia
A Flask web app to track how time is spent across business projects: relational schema, business logic and the KPIs for time efficiency.
Sep 2022
University of Modena and Reggio Emilia
Scienze e Tecnologie Informatiche, completed in July 2025.
Mar 2025
OT Consulting · Reggio Emilia
A predictive system for the success potential of new business projects: scikit-learn model, an LLM layer that writes readable reports, served with FastAPI. Analysis time went from hours to minutes.
Jun 2025 →
OT Consulting · Reggio Emilia
Python backend services and REST APIs for enterprise clients, full-stack features in Vue.js, SQLAlchemy and MySQL, ML and AI integrated into client projects, and a legacy app migrated to Google Cloud.
Dec 2026
University of Modena and Reggio Emilia
Data Engineering & Analytics track, started in September 2025. Graduation expected in December 2026.
Six projects, each with its evaluation and its limits written down. The numbers come from the repositories, not from memory.
A question-answering system over the AI Act, built on a law that changed after the model's knowledge cutoff, so the value of retrieval can be counted instead of assumed.
An end-to-end clinical ML pipeline, plus an experiment asking whether the validation protocol matters as much as the model. It doesn't, and the negative result is the most useful thing the project produced.
A computer-vision pipeline for football, from detection and tracking to team assignment, pitch calibration and a bird's-eye view, with each stage measured against hand-made ground truth.
A university project on data integration. When clinical codes are slightly wrong, exact matching misses columns that clearly belong together. A similarity built on the SNOMED hierarchy recovers them.
A parking digital twin. A webcam over a paper mock-up updates each slot's state in real time, and a Telegram bot, GPS-based barrier logic and B2B federation are built on top.
Why do employees leave? Feature engineering on badge logs, six classifiers compared, and SHAP to turn the model into actions for HR.
An animated view of a RAG pipeline, from query to chunks, scores, reranking and answer, with a live backend and a replay mode to step through a run.
Demo link coming soon
University project · Schema matching & data integration
Hospitals store the same kinds of information in different databases, with different column names. To combine them you first need to know which column in one corresponds to which in the other. That is schema matching. A common way to do it is to look at the values: if two columns contain the same things, they probably mean the same thing.
Clinical data makes this harder. Diagnoses are stored as codes from huge medical vocabularies, and real-world coding errors are rarely random. Someone records "chronic kidney disease, stage 5" when it was stage 4: a neighbouring concept, not a typo. To an exact comparison those two are as different as "kidney disease" and "broken arm".
The project treats the vocabulary as what it is, a hierarchy. Two codes are similar if they are close in that tree. I built a way to generate realistic near-miss errors on OMOP-CDM data and a matcher that compares values by meaning instead of by spelling, then measured both against exact matching on the Valentine benchmark framework.
How I approach a data or ML problem. These aren't rules I read somewhere: each one comes from a mistake I made or caught in my own projects.
Pick the metric from the decision it supports, before touching a model. With 11% positives, accuracy rewards doing nothing, so in the readmission project the metric is AUC-PR and the threshold comes from expected cost.
The expensive bugs don't crash. A "None" that means "test not ordered", patients who cannot be readmitted, a parser that quietly returns 114 units instead of 586. I look for them before I train anything.
A simple baseline tells you what the clever part is worth. In the RAG project it showed that retrieval and final accuracy come apart, which I would have missed otherwise.
Paired comparisons, confidence intervals, held-out probes for thresholds. If a design can't answer the question, change the design. That's how "the protocol matters" turned into a rejected hypothesis.
Every project ends with what it can't do. A recruiter reading a negative MOTA deserves the explanation in the same place as the number.
I use Claude Code every day as a pair programmer: to read unfamiliar code, scaffold the boring parts, write tests and challenge my reasoning. I keep the decisions: what to build, what to measure and when something is wrong.
Concretely, that means small reviewed diffs instead of big generated dumps, asking for the failure cases explicitly, and verifying numbers against the source rather than trusting a summary. This site was built that way too, with every figure checked against the repositories.
What I use in my projects and at work.
Open to conversations about AI engineering and data/ML roles, research collaborations, or a project you'd like a second pair of eyes on.