LOADING

Evaluates three routing strategies — frozen zero-shot prompting, LoRA fine-tuning, and DistilBERT classification — for automatic assignment of customer-support emails across five departments. LoRA reaches 70.4% accuracy with under 1% trainable parameters; DistilBERT peaks at 76.6%.

TF-IDF vectorisation and cosine similarity pipeline for large-scale document matching across ~10,800 Reuters news articles. Accepts raw text or file input and returns ranked results with configurable percentile thresholds — no stopword removal, operating on raw token distributions.

Five-way Nutri-Score classification (A–E) trained on 250,000 Open Food Facts products. Benchmarks Logistic Regression, KNN, SVM, Random Forest, and XGBoost across a full ML pipeline — imputation, outlier removal, PCA, stratified splits. SVM tops validation at 83.5% accuracy.

Production-ready satisfiability solver implementing congruence closure for quantifier-free first-order theories T_E, T_cons, and T_A. Follows Bradley & Manna (2007) with largest-ccpar heuristic optimisation and full SMT-LIB 2.0 support. 571 JUnit tests, all passing.

RL agents trained in the CityLearn environment for smart building energy management and real-time grid demand response. Explores actor-critic and Q-learning policies for multi-zone HVAC scheduling under dynamic electricity pricing.

Collection of bioinformatics algorithms covering pairwise sequence alignment, phylogenetic tree construction, and genomic data processing pipelines. Developed as coursework for the BSc Bioinformatics programme at the University of Verona.