Mattia Curri

GPA: 4.0/4.0, entirely in English

110 cum laude/110, GPA: 3.8/4.0

Thesis: Thesis: Supervised Learning for Semantic Segmentation of Aerial Images — 3-month research internship at CILab Lab under Prof. Gennaro Vessio, comparing CNN and Transformer architectures on the CVPR 2024 Agriculture-Vision Challenge (~50k images).

Fine-tuned Italian T5 with length-constrained generation and CSP-based grid solver. — EVALITA 2026, CEUR-WS Vol-4195

Explored whether Grad-CAM explanations improve knowledge distillation from Qwen2.5-VL-3B into a ViT student. Compared 7 distillation variants — global MSE baseline vs explanation-weighted losses with mean/attention/cross-attention probes. Built a blind-comparison arena (React/Express) with 3 local VLM judges for qualitative evaluation on 500 Mini-ImageNet images, plus BERTScore and LLM-as-a-judge automated metrics.

Production-grade multi-label classification system categorizing code comments across Java, Python, and Pharo. Built end-to-end MLOps pipeline: DVC data versioning → CatBoost + SetFit dual-model training → MLflow experiment tracking → containerized FastAPI deployment with Gradio frontend. Prometheus/Grafana monitoring, sub-2s inference latency, 92.7% test coverage, published on HuggingFace Spaces with automated CI/CD.

Forecasted future social connections in the Gab network using Evolving Graph Convolutional Networks (EvolveGCN). Integrated 768-dim BERT embeddings from user posts as node features, trained incrementally across 6 temporal snapshots. Evaluated 4 experiment configurations varying learning rate and negative sampling, with GPU-accelerated metrics (cuDF/cuGraph) and t-SNE/UMAP embedding analysis.

Predicted corporate credit ratings (4 consolidated risk classes from 22 original categories) using RandomForest, XGBoost, and LightGBM. Tackled severe class imbalance with SMOTE, ADASYN, SMOTETomek, and SMOTEENN. Built a Bayesian Network with pgmpy for probabilistic inference on missing data. Achieved 86% balanced accuracy, evaluated with Cohen's Kappa and Geometric Mean.