Curriculum Vitae
Research Interests
Mechanistic interpretability of Large Language Models: the geometry of internal representations, causal mediation of concepts (activation patching, directional ablation), and model safety, reliability, and alignment.
Education
Laurea Magistrale in Informatica (LM-18)
Università degli Studi di Milano
2024 - present
- Weighted average (exams taken): ~29.1/30
- M.Sc. thesis (starting 2026) on mechanistic interpretability of LLMs — ISLab, advisor Prof. Alfio Ferrara, in collaboration with Elisabetta Rocchetti (PhD)
- Selected coursework (grade): Natural Language Processing (30 cum laude), Distributed and Pervasive Systems (30 cum laude), Security of Service-Oriented Architectures (30 cum laude), Audio Pattern Recognition (30), Biometrics (30), Methods for Image Processing (28)
Laurea in Informatica (L-31)
Università degli Studi di Milano
2020 - 2024
- Graduated with 110/110 cum laude (weighted average 28.55/30)
- Advisor: Prof. Carlo Maria Bellettini
- Thesis: “Sviluppo di uno strumento di supporto all’associazione Recup, che recupera e distribuisce a fini sociali le donazioni del mercato ortofrutticolo di Milano”
- Relevant coursework: Software Engineering, Algorithms, Artificial Intelligence, Computer Networks
Diploma di Maturità - Perito Informatico
Istituto Tecnico Tecnologico “G. e M. Montani”
2015 - 2020
- Graduated with 100/100 cum laude
- Focus on Programming and Computer Networks
Research
The Multidimensional Geometry of Truth in LLMs — mechanistic interpretability
Research project (sole author) — basis of the M.Sc. thesis, ISLab
2025 - present
- Extended the Linear Representation Hypothesis from a single linear direction to a multi-dimensional concept cone for propositional truth
- Developed two optimization methods: TDO (gradient-based refinement of a Difference-in-Means probe under causal axioms and a KL-retention loss) and TCO (orthonormal k-D cone with Monte-Carlo interior sampling)
- Across 6 instruction-tuned models (Qwen-2.5, Gemma-2, Llama-3.1), showed a single direction suffices for smaller models but fails for larger ones, while a 2-D cone restores both causal effectiveness (Answer Switching Rate) and surgicality (low KL), with a basis nearly orthogonal to the linear probe
- Techniques: activation patching, directional ablation/addition,
nnsight; single-GPU (A40) - GitHub Repo
The Legend Challenge: Embedding Ethical Compliance into LLMs — LLM alignment / AI security
Research project with F. Bylyshi and H. El-Khazri (equal contribution) — SOASEC Legend Challenge
2026
- First empirical test of the Sargsyan–Damiani hypothesis: whether fine-tuning an LLM on narrative exemplars of compliant behaviour (legends) embeds regulatory ethics better than fine-tuning on the regulatory text, at matched backbone (
gpt-4o) and corpora - Built GenderEqGLUE, a 5-task compliance-reasoning benchmark adapted from GLUE/SuperGLUE, and an interpretability analysis via structured occlusion over API log-probabilities
- Found the two regimes teach complementary competences; led to a thesis offer at the SESAR Lab (AI security)
What Does a Heart Sound Classifier Actually Learn? — audio ML for medical diagnosis
Research project (sole author) — Audio Pattern Recognition
2026
- Recording-level phonocardiogram classification pipeline on PhysioNet/CinC 2016 comparing 3 feature representations (MFCC, log-Mel, perceptual wavelet packet) and 3 classifiers (SVM, Random Forest, CNN)
- Paired each performance claim with a falsification test (k-means geometry, per-site breakdown, leave-one-site-out, SHAP/Grad-CAM); the aggregate metric (0.866) largely reflects a recording-site shortcut, with diagnosis collapsing from 0.84 to 0.46 under leave-one-site-out
- Led to a thesis offer on a medical audio ML project
- GitHub Repo
Awards & Honors
- BOOST ‘24 — Selected for the Bologna Orthogonal Summer Term, a summer school for outstanding B.Sc. and M.Sc. Computer Science students,
Aug 2024 - Futuro Annunciato ‘24 — Selected for the event, for outstanding students,
Jul 2024 - Nomination for the “Alfieri del Lavoro” Award — Presidency of the Italian Republic,
2020 - Academic scholarship (full tuition + housing) — University of Milan,
2020 - 2024 - Merit scholarship — University of Milan,
2020 - 2023
Professional Experience
Lead Developer / Consultant
Recup APS — University of Milan
Jan 2024 - present
- Started as a B.Sc. thesis; continued as a consultancy for RECUP APS
- Designed and developed a web platform for donation tracking (React + TypeScript, Supabase DB/Auth/Storage, Netlify Functions)
- Lead developer in the post-academic phase, coordinating iterations and new features
Full Stack Developer Consultant
Value Process SRL
Jun 2022 - Oct 2022
- Developed microservices architecture using Spring Boot and React
- Led migration of legacy system to cloud-native architecture
Full Stack Developer Intern
PC & Informatica SRL
Apr 2019 - Jun 2019
- Created responsive front-end interfaces
- Optimized SQL queries improving performance by 40%
- Technologies: PHP, JavaScript, MySQL
Technical Skills
Mechanistic Interpretability & XAI
Activation patching, directional ablation/addition, Difference-in-Means probes, linear representation analysis, SHAP, Grad-CAM
Machine Learning & LLMs
PyTorch, scikit-learn, nnsight, LangChain, LangGraph, Retrieval-Augmented Generation (RAG), NL2SQL, fine-tuning, prompt engineering
Languages
Python, Java, JavaScript (ES6+), TypeScript, Go, SQL, C, OCaml, Scala, Erlang
Frameworks & Libraries
Spring Boot, React, Redux, Node.js, Flask
Tools & Technologies
- Cloud: Netlify, Supabase, Docker
- Databases: PostgreSQL, MySQL
- DevOps: Jenkins, Git, GitHub Actions
- Other: Linux
Projects
RecupGPT — Intelligent Natural-Language Querying
Retrieval-Augmented Generation over a food-surplus database
- Built with LangChain and LangGraph, integrating NL2SQL translation with query rewriting and routing
- Natural-language access to a Supabase database with thousands of records
- Cloud-native deployment (React, Supabase, Netlify Functions)
- GitHub Repo
Extended Ricart–Agrawala — Distributed Mutual Exclusion
Distributed and Pervasive Systems (30 cum laude)
- Implemented and extended the Ricart–Agrawala algorithm, resolving correctness issues arising from removing the Lamport-clock assumptions
- Developed test-first (TDD); code under a 1-year embargo
Non-Profit Information System (B.Sc. Thesis Project)
Full-stack web application for NGO operations
- React + TypeScript frontend and Netlify Functions backend
- Real-time data synchronization using Supabase
- Custom multi-role access control (RBAC)
- Features: donation tracking, volunteer management, reporting dashboards
- In use by volunteers and coordinators in the markets of Milan and Rome
Languages
- Italian: Native
- Arabic: Native
- English: Professional Proficiency (C1)