Research
Academic work from my MS at Università degli Studi di Milano — spanning audio AI, ensemble learning, and computer vision.
Includes a master's thesis, two course projects, and a literature review. Framed honestly — not inflated.
All 4 works
Study of Random Forest and Its Variants
Master's thesis — ensemble learning, hyperparameter tuning, and imbalanced classification
—Compared 4 RF variants across credit card fraud, breast cancer, and heart disease datasets
—BalancedRandomForest improved minority-class recall from 71% → 92% on fraud detection
—Identified class_weight as the single most impactful hyperparameter for imbalanced classification
—Full Python/Scikit-learn implementation with stratified 5-fold cross-validation
Urban Sound Classification with Convolutional Neural Networks
Course project — 82% accuracy on UrbanSound8K using CNN + MFCC feature extraction
—82% test accuracy on UrbanSound8K — competitive with state-of-the-art at submission
—7-layer Conv1D CNN processing MFCC features, outperforming KNN, Decision Tree, Random Forest baselines
—Supervised by Prof. Nicolò Cesa-Bianchi — one of Europe's most cited ML researchers
—Foundation for current production TTS/STT pipelines at Edza.ai
Bird Voice Recognition Using CNN and MFCC
Academic paper — multi-class bird species identification from field audio recordings
—88-class bird species identification from field recordings — CNN achieves 40% vs 1.1% random chance
—Extracted 6 spectral features: MFCC, Mel-spectrogram, Chromagram, Spectral Centroid, Roll-off
—Supervised by Prof. Stavros Ntalampiras — specialist in bioacoustics and audio ML
—Identified data scarcity as the core bottleneck, motivating transfer learning future work
Face Recognition in Neural Networks: A Literature Review
Synthesis of 15+ IEEE papers covering CNN architectures, GANs, and heterogeneous face recognition
—Synthesis of 15+ IEEE papers spanning FaceNet, ArcFace, DeepFace, and GAN-based approaches
—Covers 4 research threads: standard FR, heterogeneous (cross-modal) FR, super-resolution, multi-task learning
—Identifies loss function design (ArcFace margin loss) as more impactful than architecture choice
—Bridges to audio AI: connects face embedding techniques to speaker verification literature
All work completed during the MS programme at Università degli Studi di Milano, Italy — under the supervision of faculty in the Department of Computer Science.