Mattia Curri

• Course entirely taught in English
• Relevant courses: Numerical Methods, Database II, Formal Methods, Fundamentals of Artificial Intelligence, Natural Language Processing, Software Engineering for AI-Enabled Systems

• Thesis in Computational Intelligence: Supervised Learning Techniques for Semantic Segmentation of Aerial Images of Agricultural Fields
  • 3 Months research internship at the Cilab Lab under the supervision of Prof. Gennaro Vessio.
  • Comparison of Convolutional and Transformer architectures for the Agriculture-Vision Challenge at CVPR 2024, training models on ~50k images.
• Relevant courses: Computational Intelligence, Algorithms and Data Structures, Databases, Software Engineering, Knowledge Engineering, Information Retrieval, Human-Computer Interaction, Computer Networks.

• Architected a production-grade multi-label classification system categorizing code comments across Java, Python, and Pharo
• Engineered end-to-end ML pipeline: data versioning (DVC) → automated preprocessing → dual-model training (CatBoost + SetFit) → experiment tracking (MLflow) → containerized deployment
• Implemented comprehensive data quality framework improving dataset integrity by 12%
• Deployed full-stack solution: FastAPI backend with async processing → Gradio frontend → Prometheus/Grafana monitoring, with sub-2s inference latency
• Achieved 92.7% test coverage across unit, integration, and behavioral test suites
• Published live demo on Hugging Face Spaces with automated CI/CD

• Developed a comprehensive Retrieval-Augmented Generation system analyzing ~20 legal documents (up to 200 pages each)
• Built end-to-end pipeline: web scraping → PDF processing → knowledge graph extraction with Gemini → semantic indexing → RAG inference
• Implemented knowledge graph normalization using DBSCAN clustering and semantic embeddings to reduce entity redundancy
• Created custom evaluation framework with Context Faithfulness, Context Precision/Recall, and Answer Accuracy metrics
• Integrated multiple retrieval strategies: standard, multi-query, entity extraction, and Neo4j random walks

• Developed a temporal link prediction system using Evolving Graph Convolutional Networks (E-GCN) to forecast future social connections in the Gab network
• Integrated BERT embeddings (768-dimensional) to capture semantic content from user posts and enhance prediction accuracy
• Engineered a multi-snapshot temporal pipeline handling 6 historical time steps with incremental training and fine-tuning
• Achieved MAP=0.87, Macro F1=0.86, AUC-ROC=0.80 on temporal link prediction tasks

• Predicted ratings assigned by rating agencies to companies
• Handled unbalanced dataset using Class Weights, SMOTE, ADASYN, SMOTETomek, SMOTEENN
• Implemented a Bayesian Network for probabilistic queries, with Balanced Accuracy of 0.59
• Achieved up to 86% Balanced Accuracy using XGBoost and Random Forest

A. Porcelli, F. Di Gravina, E. Fontana, M. Curri, F. D. Di Gregorio. “FFT-UniBa at Cruciverb-IT: Special Length Tokens and CSP for Italian Crossword Solving.” In EVALITA 2026, Pre-print: link

• Fine-tuning of Italian T5 with length-constrained generation and CSP-based grid solver; MRR 0.63, 34% full grid accuracy.

• Achieved 2nd place in the first stage, out of 20 teams.
• Used open data to create a solution for the city of Bari, focused on logistics and intelligent traffic management.

• Anomaly Detection task in Hadoop Distributed File System Logs.
• Secured 2nd place out of 15 teams.

• Winning team of the AI2B Hackathon organized by University of Bari and AI2B, focused on AI and cybersecurity.
• Obtained €500 winning first place out of 10 teams.

• Qualified to the national stage by ranking in the top 6 of the local venue, reaching 21st place out of 43 teams nationally.
• Qualified for the local stage by being in the top 20 of the selection.