Computer science graduate with a BSc (2023) and MSc (2025) from Università della Svizzera Italiana (USI), early stage research experience in machine learning, federated learning, and human activity recognition, well rounded technical skills across the CS spectrum, and experience in leading development projects.
My work resulted in academic publications on federated learning, human activity recognition, and privacy-aware machine learning for mobile and wearable devices.
Follow-up of my Bachelor thesis. Proposes a missing-modality-agnostic federated learning for human activity recognition on head-worn wearables; robust to missing sensors.
Published in: IEEE Pervasive Computing, 23(4): 40-49
Proposes a federated learning approach for cognitive workload estimation using wearable sensors, achieving performance comparable to centralized models while preserving data privacy.
Published in: Proc. International Conference on Mobile and Ubiquitous Multimedia (MUM '23)
Academic theses showcasing my research in machine learning and federated learning and my development skills on local and distributed systems.
Combining Contrastive and Pretext-Based Learning for Effective Encoders from Unlabeled and Federated Data in Human Activity Recognition and Beyond
Master's thesis on Multi-Task Self-Supervised Learning for label-efficient learning. Modular PyTorch framework combining contrastive + pretext tasks with dynamic loss weighting, and centralized/federated training (HAR/STL-10) to learn compact, robust representations.
Institution: Università della Svizzera Italiana (USI)USI Verify diploma
Bachelor thesis on federated learning for multimodal sensor data, focusing on privacy-preserving and robust techniques to handle missing modalities.
Institution: Università della Svizzera Italiana (USI)USI Verify diploma
Selected personal and academic projects covering as different subjects as machine learning, web development, mobile apps, robotics, visual computing and computational fabrication. Projects include individual and collaborative work demonstrating a range of skills and technologies. For most of the collaborative projects, I acted as team leader.
Code for my Master's thesis on Multi-Task Self-Supervised Learning for label-efficient learning. Modular PyTorch framework combining contrastive + pretext tasks with dynamic loss weighting, and centralized/federated training (HAR/STL-10) to learn compact, robust representations.
Retrieval-augmented chatbot that answers Swiss health insurance questions using a vector database and large language model. The system integrates semantic embeddings with FAISS vector storage to index insurance documents, accessible through Streamlit web interface, Jupyter notebooks, and voice conversation with speech-to-text transcription and text-to-speech synthesis. User queries retrieve contextually relevant passages through semantic search, fed into a LangChain prompt template that grounds LLM responses with cited document sources and relevance scores.
Arduino-based sensor network that tracks noise pollution and crowd density across urban spaces. The system consists of several sensor nodes that collect data using camera and sound sensors to monitor environmental conditions and transmit the information to a central server together with GPS coordinates and timestamp for real-time mapping and analysis. The images captured by the camera sensors are processed using computer vision techniques to estimate crowd and traffic density.
Android navigation app for USI campus with QR code scanning, voice commands, offline maps, and crowd-sourced map extension. Features real-time navigation, camera integration, and Gemini API assistance.
A computational fabrication tool that stabilizes 3D shapes by voxelizing an input OBJ mesh and carving interior voxels to shift the center of mass over the support region. It voxelizes the model using multi-direction ray casting with Fibonacci sphere sampling and jitter, extracts the lowest voxel layer as a base, computes its convex hull via Graham scan, and repeatedly recomputes the barycenter while removing interior voxels farthest along a semiplane opposite the hull to migrate mass. If carving fails to achieve balance (barycenter inside projected convex hull), it dilates the base outward in the barycenter direction and rechecks stability, then exports the final voxel configuration as OBJ artifacts.
Thymio robot simulation that completes a maze using flood-fill path planning and sensor-driven navigation. A dynamically sized grid stores wall bitmasks and distance costs, incrementally updated from either camera-based occupancy extraction or proximity sensors before flood-fill recomputes goal-oriented distances. Mirroring Micromouse phases, the controller (sense → update → propagate → select) first reaches the goal, then returns to start while exploring more unknown cells, then performs an optimized speed run using the cached minimal path without further full-map recomputation.
A high-performance C++ raytracer awarded first place in the USI rendering competition. The system implements physically based lighting and materials, combining tile-based OpenMP parallelism with an AABB tree acceleration structure for efficient triangle mesh intersection. It supports roughness-driven BRDFs, multi-map texturing (diffuse, normal, roughness, displacement), procedural texture generation, and reflective and refractive materials. The hierarchical acceleration pipeline and extensive parallelization enable fast rendering of complex scenes while maintaining accurate recursive reflection and refraction.
Search and discovery platform with custom ranking and filtering to explore content creators, built on a Python Scrapy crawler that aggregates creator data from Patreon, Ko-fi, and SubscribeStar into a Solr search index. The Vue.js frontend queries a Node.js backend API to deliver faceted search results ranked by custom scoring algorithms, enabling users to discover creators through advanced filtering by category, engagement metrics, and performance indicators. Results are personalized based on user preferences and interaction history.
Led a 14-person team to build an educational web app that teaches programming fundamentals through interactive Blockly exercises and real-time 3D animations. Learners create programs using Blockly's drag-and-drop interface, which are first validated and executed on the backend to ensure correctness and safety. The resulting instruction sequence is then animated by an on-screen robot using a Three.js pipeline that visually mirrors the program's logic.
Collaborative text editor web application featuring real-time synchronization and sharing tools. Built with Express.js and Node.js on the backend, it uses ProseMirror with operational transformation to handle concurrent editing while keeping documents in sync. Socket.IO manages real-time communication between clients, MongoDB stores documents and user data, and Passport.js handles authentication. When users edit, changes flow through WebSockets to the server, get applied to the document, and broadcast instantly to other collaborators.
Java-based spreadsheet application with formula evaluation, validation, and persistent storage. The
spreadsheet stores cells in a HashMap with formulas represented as Abstract Syntax Trees. When a cell
is updated, dependent cells are automatically marked for recalculation. Supported operations include
basic arithmetic, trigonometric functions (SIN, COS, TAN), and range operations (SUM, AVERAGE, MIN,
MAX, COUNT). The system detects circular references during evaluation. Data can be saved in a custom
.cellarium format or exported as CSV, with full undo/redo support through state
snapshots. Both text-based and graphical interfaces are available.