Projects

A selection of things I’ve built and studied.

ML · Data · Bioinformatics • Last updated: February 2026

Featured: Bachelor’s thesis

Unveiling Protein Connections through Graph Neural Networks and ProtT5 Embeddings
A lightweight link-prediction pipeline for protein–protein interactions combining ProtT5 sequence embeddings with graph context (STRING v12). Includes a NOTCH2-focused case study.

Task: predict missing PPIs using an encoder–decoder GNN with learned node representations.
Data: high-confidence STRING graph + ProtT5 embeddings as node features.
Result highlight: strong separability and excellent early precision (top-ranked predictions).

Read thesis (PDF) Slides (PDF)

If you want details about setup/metrics, the full writeup is in the thesis PDF.

UAV Anomaly Detection & Localization

Detection model created for the Sapienza Flight Team to detect and localize targets in high-resolution aerial imagery using reconstruction-error based models and post-processing (masks → bounding boxes → classification-ready crops).

Focus: robust localization under “static background + rare objects” assumptions.
Work: AE/VAE variants, denoising strategies, thresholding + connected components for bboxes.
Outcome: fast candidate generation that can be cheaper than running a detector everywhere.

EU Elections 2019 Data Analysis

Exploratory analysis of the 2019 European elections: cleaning results, testing hypotheses, and producing visual summaries.

Work: wrangling messy tables, consistency checks, reproducible notebooks.
Output: plots + narrative conclusions based on quantitative patterns.

Presentation (PDF) , a review and expansion of this project is planned for the future

Lorenz Attractor Forecasting (Multi-output LSTM)

Small toy notebook forecasting the Lorenz system with an LSTM predicting coordinates and derivatives, exploring stability in rollouts and training dynamics.

Work: multi-head outputs, delta/residual ideas, sensitivity to derivative scaling.
Outcome: learned which formulations remain stable in multi-step prediction.

Notebook (ipynb) • Rendered (HTML)

Synthetic Dataset Generation (Blender)

Small pipeline for generating labeled synthetic images for the SUAS-2025 competition to which the Sapienza Flight Team partecipated, with domain randomization for YOLO training.

Work: scene randomization, lighting changes, exporting labels.
Why: quick iteration when real data is expensive or limited.

Tip: keep images ~1200px wide for fast loading.

Small Web / API Projects

Lightweight backend/API work and small apps to wrap functionality into something usable and testable.

Work: API schemas, endpoints, and clean integration with client logic.
Stack: OpenAPI + SQL fundamentals + HTML/CSS/JS for simple UIs.

WasaText • Mood tracker