ML Engineer at Intesa Sanpaolo's AI Delivery Center, finishing my M.Sc. in Artificial Intelligence & Data Analytics at Politecnico di Torino. My thesis, developed as a graduate intern at Data Reply, focuses on production-oriented GenAI for metadata management.
Most of my work sits between LLMs and the plumbing around them: RAG pipelines with LangGraph, fine-tuning with PEFT/LoRA, retrieval quality (hybrid search, reranking), and the boring-but-important parts like evaluation and quantization for hardware that isn't a datacenter.
SemanticMesh — M.Sc. thesis project LangGraph multi-agent system for data governance: it aligns business documents with relational schemas in a Neo4j knowledge graph. A builder pipeline handles triplet extraction, entity resolution, actor-critic validation and Cypher healing; a query graph combines dense + BM25 + graph traversal with cross-encoder reranking and hallucination grading. Evaluated on 7 datasets, 111 tables, 210 questions: 210/210 grounded answers, zero hallucinations.
Multi-Turn Self-Corrective RAG — SemEval 2026 Task 8 Self-CRAG system with LangGraph for grounded multi-turn answers. Hybrid retrieval with parent-child chunking, BGE-M3 embeddings and cross-encoder reranking, feeding a 4-bit quantized Llama 3.1 8B for fully offline inference on consumer hardware.
SM-SIP — multilingual abstractive summarization Controllable summarization with semantic supervision. IT/EN pipeline on Llama and Qwen, with token-classification modules that flag hallucination-prone spans before generation. Package and LoRA adapters (PEFT) on the Hugging Face Hub.
Concept Discovery for Medical VLMs — explainability Unsupervised concept discovery in medical vision-language models via Sparse Autoencoders, with a concept-naming module and an LLM-as-a-judge setup to measure how faithful the concepts actually are.
DYLEM-GRID — gesture recognition BiLSTM with attention and an encoder-only Transformer for dynamic hand-gesture classification from Leap Motion time series. The dataset (400 gestures, 100 participants) is on Kaggle.
Core: Python, PyTorch, Hugging Face, LangChain / LangGraph, Pandas, Scikit-learn LLM work: RAG, fine-tuning, PEFT / LoRA, quantization, prompt engineering, hybrid retrieval and reranking Ops & tools: Docker, Git, Weights & Biases, SQL, Claude Code, Copilot
- M.Sc. Artificial Intelligence & Data Analytics, Politecnico di Torino (2024–2026, expected 110/110) Thesis: Generative AI and Foundation Models for Automated Data Engineering and Metadata Orchestration
- B.Sc. Computer Engineering, University of Enna "Kore" (110/110 cum laude) Thesis: Recurrent Neural Networks for Dynamic Gesture Recognition
Cambridge English C1 Advanced (2024).
