I hold an MSc in Artificial Intelligence from the University of Bologna – Alma Mater Studiorum, where I graduated in 2026 with a thesis on the robustness of optimization-based decision support systems for urban green planning. I'm passionate about machine learning, Computer Vision, NLP and decision-making systems, and I like projects where theory turns into something people can actually use. When I need a tool that doesn't exist (or doesn't work the way I want), I build it myself.

• TALEA DSS Robustness: Robustness assessment of an optimization-based decision support system for placing modular green interventions in Bologna (TALEA Green Cells, EU Urban Initiative). Extends the existing MiniZinc placement pipeline with a perturbation framework (Gaussian/multiplicative noise, mislabelling, missingness, geometric and aggregation perturbations at 5 severity levels), spatial and quantitative robustness metrics (Jaccard, buffered IoU, utility and allocation ratios) and a cross-model comparison of average stability and worst-case fragility. Fully containerized with Docker.

Computer Vision

• Open-World Semantic Segmentation: Anomaly segmentation of unexpected objects on roads (StreetHazards). A SegFormer-B0 fine-tuned with weighted CE + Dice and rare-class cropping, combined with post-hoc OOD scoring (MSP, Max Logit, Energy) and Boundary Suppression. The best setup (Energy + BS) reaches ~1.8× the AUPR of the MSP baseline.
• Sea Surface Temperature Reconstruction: A U-Net with dilated convolutions that reconstructs sea surface temperature under cloud occlusion in MODIS satellite images of the North Adriatic, trained with a masked RMSE loss on artificially clouded samples. Cuts reconstruction error by ~25% compared to a tuned climatological baseline.

NLP & Large Language Models

• Audience Reaction: An AI-powered audience intelligence platform that turns the comments under a YouTube video into an evidence-backed report: sentiment toward creator, content and sponsor, recurring topics, and the comments that support each finding. A multi-stage LLM pipeline (Jev classifier, Claude Sonnet topic discovery and consolidation, topic assignment) developed with a rigorous evaluation process: hand-built synthetic datasets with gold labels, leakage audits, held-out sets and pre-registered experiments. Built with TypeScript and Next.js.
• Multimodal Summarization: Multimodal (text + audio) summarization with Gemma and Qwen. Includes inference with greedy, beam search and temperature sampling, fine-tuning of Gemma on XSum and SAMSum, and evaluation with ROUGE plus qualitative analysis.
• GenSPP × ERASER Benchmark: Training GenSPP (Genetic Select-then-Predict Pipeline) on ERASER datasets (Movies, e-SNLI) and evaluating the extracted rationales with the official ERASER metrics.
• Sexism Detection: Sexism detection using RNNs, Transformers and prompting techniques.

Fairness & Responsible AI

• Fair Candidate Screening System: Fairness-aware ML for candidate screening, comparing bias-mitigation techniques such as reweighing, FaUCI and learned fair representations while preserving predictive performance.

Decision Making, Optimization & Game AI

• TablutAI: An AI agent that plays Tablut, the strategic board game inspired by the Viking game Hnefatafl.
• CDMO Project: Combinatorial Decision Making and Optimization project developed at UniBo.

• Dark Enigma (iOS, built with a team of collaborators): A mystery game where you solve dark lateral-thinking cases by questioning an AI narrator.
• Notion Expense (Android, free & open source): Log an expense into your Notion finance tracker in a few taps. Quick Add card (launchable with a double tap on the back of Samsung phones via Good Lock), offline queue with reliable background sync (WorkManager), and a dashboard with monthly budget, charts and filters. Built with Kotlin and Jetpack Compose. I built it because I needed it, and I use it every day.

• Languages: Python, Java, Kotlin, TypeScript
• ML / DL: PyTorch, PyTorch Lightning, Hugging Face Transformers, TensorFlow / Keras, LLM fine-tuning
• LLM engineering: multi-stage LLM pipelines, prompt and contract design, LLM evaluation and benchmarking
• Optimization: MiniZinc, combinatorial optimization
• Focus areas: NLP, Computer Vision, Multimodal Learning, Optimization, Robustness, Fairness in ML
• Apps & Web: Next.js, Jetpack Compose (Android)
• Tools: Jupyter, Google Colab, Kaggle, Docker, Git