I build intelligent systems end-to-end.

Master's thesis · Politecnico di Torino

LangGraph-orchestrated multi-agent system aligning business documents with relational schemas in a Neo4j knowledge graph. Two pipelines: a builder (triplet extraction, entity resolution, actor-critic validation, Cypher healing) and a query graph (dense + BM25 + graph traversal, cross-encoder rerank, hallucination grading). Evaluated on 7 datasets · 111 tables · 210 questions: 210/210 grounded answers, zero hallucinations, 4.31/5 mean score, 21 ablation studies.

SemEval 2026 — Task 8

Self-Corrective RAG with LangGraph for grounded answers across multi-turn dialogues. Hybrid retrieval (Parent-Child chunking, BGE-M3 embeddings, Cross-Encoder reranking) plus 4-bit NF4 quantization of Llama 3.1 8B for offline inference on consumer hardware.

Abstractive Summarization · IT/EN

Controllable abstractive summarization with semantic supervision. Multilingual (IT/EN) pipeline integrating Llama and Qwen with token-classification modules to detect and mitigate hallucinations. LoRA adapters (PEFT) published on the Hugging Face Hub.

Medical Vision-Language Models

Unsupervised concept discovery in medical VLMs via Sparse Autoencoders. A concept-naming module paired with an LLM-as-a-judge system to quantify concept faithfulness.

Dataset: 400 gestures · 100 participants

BiLSTM with attention and an encoder-only Transformer (PyTorch) for dynamic hand-gesture classification. Optimized high-dimensional feature extraction on the DYLEM-GRID dataset, also published on Kaggle.