The full record.

BSc in Artificial Intelligence

Education — Università di Pavia / Università degli Studi di Milano / Università degli Studi di Milano-Bicocca · Pavia / Milan, Italy

Inter-university Bachelor's programme in Artificial Intelligence, jointly run by three Italian universities and launched the year of enrolment — graduated in its inaugural cohort.

• Core coursework spanning machine learning, deep learning, statistics, algorithms, programming and databases, alongside knowledge representation & reasoning, computational logic and cognitive psychology.
• Graduated on 26/09/2024.
• Bachelor's thesis, "Neural Network Tool for Hyperparameter Optimization for Embedded Systems," developed on-site during the STMicroelectronics internship.

Industrial Robotics Trainer & Certification Assessor

Work — Comau × Pearson (2022–2023) → Fanuc × Sanoma (2023–present) · Northern Italy

Delivers technical training and administers official industry certifications on industrial communication protocols, robot programming and Industry 4.0 to mixed audiences of engineers, educators and students.

• Produces structured course materials and assessment criteria and leads "Train the Trainer" sessions.
• Translates dense technical content into reproducible curricula for audiences ranging from vocational students to practising engineers.

Science Museum Guide

Work — Leonardo da Vinci National Museum of Science and Technology · Milan, Italy

Communicates complex technical and historical concepts to diverse public audiences across the museum's aviation, transport and historical-technology exhibitions.

AI Engineer Intern — Bachelor's Thesis

Work — STMicroelectronics · Agrate Brianza, Milan, Italy

Designed and implemented an automated hyperparameter-optimisation framework, extending the open-source Keras Tuner library, for neural networks deployed on microcontroller-class hardware — targeting an industrial audio anomaly-detection use case for predictive maintenance.

• Engineered the search procedure to jointly evaluate candidate architectures against accuracy and strict embedded constraints — ROM footprint, RAM usage and Multiply-Accumulate operations (MACCs).
• Validated the framework's ROM/RAM/MACC estimators against STM32Cube.AI ground-truth measurements across three MLPerf Tiny reference models (anomaly detection, image classification, keyword spotting).
• Replaced an initial "poisoning" strategy for oversized trials with a cleaner trial-skipping mechanism built on Keras' FailedTrialError API.
• Diagnosed and engineered a workaround for an upstream Keras Tuner bug that terminated an entire search whenever the first proposed configuration violated the constraints, keeping long unattended remote tuning runs alive.
• Operated inside STMicroelectronics' confidential corporate environment — remote access to internal Linux workstations via company VPN — with fully reproducible, version-controlled Conda environments and systematic experiment logs.
• Findings compiled into the Bachelor's thesis, co-supervised academically at the university and industrially at STMicroelectronics.

MSc in Computer Science & Engineering (Artificial Intelligence)

Education — Politecnico di Milano · Milan, Italy

Master's specialising in Artificial Intelligence — machine learning, deep learning, natural language processing, large language models, neural networks, distributed systems and software engineering.

• Deliberately chose a project-based path over an exam-heavy one — slower and more demanding, but it placed him in direct collaboration with professors and researchers on real problems, with Python as the working language throughout.

Tooth Segmentation Transformer (ToothFormer)

Project — Politecnico di Milano, coursework · Milan, Italy

A SegFormer-inspired Vision Transformer for conditional semantic segmentation.

• Identifies and masks individual teeth in 3D dental scans, conditioned on a given tooth ID rather than segmenting the full arch at once.

View on GitHub ↗

TV Show Recommender

Project — Politecnico di Milano, coursework · Milan, Italy

A recommendation system for TV shows using user-user and item-item collaborative filtering.

• Analyses viewing patterns and computes similarity matrices.
• Features an interactive 3D visualisation of the recommendation space.

View on GitHub ↗

WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models (Part I & II)

Project — Politecnico di Milano, exam project for "Large Language Models: Applications, Opportunities and Risks" · Milan, Italy

Two-part exam project completed with one course partner: a faithful replication and critical analysis of the WebVoyager end-to-end web-agent framework, followed by experiments to reduce execution time and cost, plus a literature review tracing the evolution of web agents.

• Part I replicates WebVoyager and evaluates modifications aimed at reducing execution time and operational cost while improving overall effectiveness, benchmarked against the reference implementation.
• Part II surveys the emerging web-agent landscape and details the ethical and safety risks of the technology.
• Seeded the later, more ambitious CUHK-Shenzhen replication that swapped in fully open-weight, locally-hosted models.

View on GitHub ↗

Demonstration of a Simple Transformer Running on the NPU of an STM32N6

Project — Politecnico di Milano, "Advanced Computer Architectures" course project · Milan, Italy

Full-stack edge-AI engineering project completed with one course partner: designed, trained and INT8-quantised a custom 4.56M-parameter encoder-only Transformer from scratch for autoregressive text generation, then deployed it on the Neural Processing Unit of an STM32N6570-DK.

• Validated the toolchain first against a Model Zoo MobileNetV2 and a hand-crafted CNN before tackling the Transformer as the project's core objective.
• Applied per-channel symmetric INT8 quantisation of both weights and activations against a held-out calibration set.
• Full pipeline: PyTorch model design, training on ~1.4M words from Project Gutenberg, C-code generation via STM32CubeIDE / X-Cube-AI, NPU configuration and on-device validation.
• On-board results: 73.14 ms/token latency, 13.67 tokens/s throughput, 257 of 407 ops accelerated on the NPU — with a quantitative CNN-vs-Transformer comparison under identical hardware constraints.
• Models trained locally on a workstation with two NVIDIA GeForce RTX 5060 Ti GPUs.

View on GitHub ↗

Exchange Student — Computer Science & Engineering

Exchange — The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen) · Shenzhen, China

A semester of AI research and advanced coursework in machine learning and engineering.

• Coursework included CIE6004 (Image Processing and Computer Vision) and CIE6006/MCE5918 (Data Analysis), alongside independent research collaborations with CUHK-Shenzhen peers.
• Continued running compute-heavy Polimi coursework and personal-research projects remotely on his own dual-RTX-5060-Ti Linux workstation back in Milan throughout the exchange.

Sen2Fire Wildfire Detection — Reproducibility and Enhancement

Project — Independent research collaboration · Shenzhen, China / Milan, Italy

A rigorous re-evaluation and enhancement of the Sen2Fire wildfire-detection benchmark, developed with an international three-person team spanning CUHK-Shenzhen and HSE University Moscow.

• Benchmarked a modified U-Net baseline against nnU-Net, SegFormer, Dual-Stream SegFormer and MixedTrans-CNN architectures on Sentinel-2 / Sentinel-5P multispectral imagery.
• Established a new state-of-the-art F1-score of 0.3675 (IoU 0.2250) with nnU-Net on all 13 spectral bands, versus the original paper's 0.2810 — challenging its conclusion that band subsets outperform the full spectral range.
• Added Test-Time Augmentation (six-view flip/rotation ensembling) and quantified its cost/benefit trade-off against the roughly six-fold increase in inference latency it incurs.

View on GitHub ↗

Deep Learning From Scratch — CIFAR-10

Project — CUHK-Shenzhen, CIE6006/MCE5918 Data Analysis coursework · Shenzhen, China

Implementation of Batch Normalization, Dropout and Convolutional Neural Networks using raw NumPy alongside a PyTorch counterpart, for the CIFAR-10 dataset.

View on GitHub ↗

Replicating and Enhancing WebVoyager

Project — CUHK-Shenzhen, CIE6004: Image Processing and Computer Vision · Shenzhen, China

A two-person CUHK-Shenzhen coursework team replicated and extended the WebVoyager framework for end-to-end AI web agents, replacing proprietary GPT-4o with locally-hosted open-source Large Multimodal Models to test a scalable, cost-free path to self-hosted web agents.

• Deployed Qwen3-VL:8b and Qwen3-VL:30b locally via Ollama and integrated DeepSeek-V3 (685B MoE) via OpenRouter as a third comparison point against GPT-4o.
• Built a fully-automated document/page-classification pipeline using locally-hosted Qwen3-VL:30b, producing a 1,101-image labelled dataset with zero manual annotation.
• Fine-tuned a ResNet-18 CNN for real-time page-state classification as a visual prior in the agent loop, cutting reasoning latency by 17%.
• Designed a novel grayscale + colour Set-of-Mark visual-grounding technique (52.4% vs. 51.1% baseline task success); replicated the GPT-4o reference at 62.7% task success, above the original paper's 59.1%.

View on GitHub ↗

Retrieval-Augmented Generation (RAG) for LLM-Based Code Generation

Project — Politecnico di Milano, "Software Engineering 2" course project · Milan, Italy (run remotely from Shenzhen, China)

A three-person course-exam project investigating whether retrieval-augmented generation and prompt engineering improve LLM-based code generation for library-centric tasks (PySCF and SEED-Emulator), using CodeLlama-7B-Instruct over a knowledge base of 7,600+ code snippets.

• Engineered and benchmarked three single-hop retrievers (lexical BM25, semantic dense via CodeBERT 768-d embeddings, hybrid via Reciprocal Rank Fusion) and two LLM-driven multi-hop pipelines (query decomposition, iterative refinement).
• Deployed the end-to-end system fully on-premises on a dual-RTX-5060-Ti Linux workstation with zero external API calls, using 4-bit NF4 quantisation (bitsandbytes) to fit CodeLlama-7B-Instruct in available VRAM.
• Result: semantic retrieval delivered a +38% relative CodeBLEU improvement over the no-retrieval baseline (0.1859 → 0.2567); multi-hop decomposition with a checklist-driven prompt reached 0.2505, rivalling the best single-hop setup.
• Found that hybrid RRF underperformed pure dense retrieval on this corpus, while minimal prompts worked best for the no-retrieval baseline — added structure without external context introduced overhead.

View on GitHub ↗

GenHack 2025 — Correcting Urban Heat Island Bias in ERA5 Climate Data

Project — École Polytechnique × BNP Paribas × Kayrros, hackathon (Team BAT_TEAM) · Remote

Three-person team hackathon: built an ML pipeline correcting systematic Urban Heat Island bias in ERA5-Land coarse-resolution (9 km) climate reanalysis data.

• Integrated four heterogeneous data sources — ERA5-Land, Sentinel-2 NDVI, ECA&D ground stations, JRC Global Surface Water — into a unified 3M+ row training dataset across Italy, Spain, France and Portugal.
• Trained and compared Random Forest vs. HistGradientBoosting regressors with a station-based split to prevent geographic leakage; engineered a land-cover segmentation step fusing Sentinel-2 NDVI with JRC water data to resolve urban/water spectral ambiguity.
• Result: reduced ERA5 raw error from 3.01°C to 1.94°C RMSE (−35.5%), MAE by −45% and corrected systematic urban bias by −93% on hold-out test stations.

View on GitHub ↗

AI Traineeship — Profile 1: Local AI Model Architecture & Infrastructure

Work — European Court of Auditors (ECA), Directorate for Information Technology, Workplace and Innovation (DIWI) · Luxembourg

Upcoming traineeship in the Innovation Atelier, building on-premises open-source AI under strict data-sovereignty and confidentiality constraints — local LLM deployment, RAG, OCR-driven document ingestion and document classification — directly extending the on-prem discipline built at STMicroelectronics and in the RAG/WebVoyager coursework.

• Applied specifically for Profile 1 rather than a general AI internship, drawing on daily practice self-hosting open-source LLMs (Llama, Qwen, DeepSeek) via Ollama on a dual-GPU Linux workstation.
• Plans to containerise the existing Ollama + RAG stack with Docker Compose and stand up a Tesseract/PaddleOCR ingestion pipeline ahead of the start date, to arrive ready to integrate rather than ramp up.
• Extends the on-device/microcontroller reasoning discipline built during the STMicroelectronics internship directly into on-premises institutional AI deployment.