Introduces a data-free preconditioning method for privacy-preserving deep learning, enabling effective model training under differential privacy constraints without requiring access to raw training data.
Read paper →CERN · UPF · Kosmico
Full-time PhD student and founder. Open source projects on the side.
Research scientist and PhD student at UPF, based at CERN in Geneva. I work on multimodal generative models and agentic systems for neurological diagnostics: federated learning, brain signals, and privacy-preserving medical AI. Co-founder of Kosmico, an AI workspace for collaborative research. MSc in Data Science and BSc in Statistics from Sapienza University of Rome.
Co-founder
Kosmico
The all-in-one AI workspace for collaborative research
Everything in one place, built for collaboration with agents and teammates. Write, cite, plan, and run projects with Kosmo, an agent that lives inside your workspace and actually does the work.
Native AI workspace
LaTeX, notebooks, PDFs, and code in one environment, with @kosmo always one message away.
Built for labs
Shared projects, role-based access, channels, and calendars. Already powering researchers at CERN and universities.
Research loop, closed
Literature review, bibliography management, deep research across academic databases, and knowledge graphs, all connected.
About
Multimodal generative models & agentic systems
My research sits between multimodal generative models and agentic systems: learning latent structure from brain signals on one side, and building agents that reason through clinical investigations with tools and protocols on the other. I also care about privacy, federated learning, and making models explainable in real clinical settings.
8
Publications
25
Citations
3
Institutions
2
Open-source projects
Multimodal Generative Models
Generative modelling of brain signals and neuroimaging. Flow-based latent learning from EEG, capturing uncertainty in cortical inference.
Agentic Systems
Tool-augmented agents for clinical neurology. Step-by-step investigation with diagnostic tools, hospital protocols, and evidence-based reasoning.
Private & Explainable AI
Federated learning and privacy-preserving training across hospitals. Explainable methods that stay interpretable when data cannot be shared.
Publications
8 papers · 25 citations
Proposes a federated architecture combining Transformers and Graph Neural Networks for brain tumor localization across multimodal MRI, with modality-level explainability while preserving patient privacy.
Read paper →Presents a semantics-aware communication fabric enabling large-scale coordination among AI agents, designed for scalable multi-agent systems with structured knowledge exchange.
Read paper →Characterizes the heterogeneity amplifier effect, where differential privacy noise disproportionately degrades explanation fidelity on heterogeneous clients, and proposes BID-CAM, a DP-aware hybrid explanation method for federated 3D medical image segmentation.
Read paper →Introduces a hierarchical encoder combining patch-level Transformers with supervoxel-level Graph Attention Networks for brain tumor localization, offering dual-scale explainability without a decoder module.
Read paper →Combines Federated Learning and Graph Neural Networks to predict stroke severity from EEG data across multiple hospitals, preserving patient privacy while achieving competitive diagnostic performance.
Read paper →Evaluates the feasibility of federated neural networks for explainable atrial fibrillation detection, addressing the challenge of early detection in asymptomatic and paroxysmal cases.
Read paper →Proposes a Graph Neural Network approach to predict stroke severity (NIHSS) from EEG recordings of 71 patients, using graph attention to reveal frequency-dependent brain reconfiguration patterns for clinical decision-making.
Read paper →Experience
Work & education
Work
Building the AI workspace for collaborative research: writing, citing, planning, and agent-assisted workflows in one place.
PhD Student, CERN
Developing generative models and agentic systems for neurological diagnostics, with a focus on multimodal brain signals, federated learning, and privacy-preserving clinical AI.
Research Scientist Intern, CERN
Graph Neural Networks for medical diagnostics, predicting stroke severity through brain connectivity analysis from EEG signals.
External AI Consultant, Sorridi
Deep Learning for dental aligner automation from 3D meshes and patient images. Led a team of five fine-tuning HuggingFace models.
Education
PhD in AI for Neuroscience, Universitat Pompeu Fabra (UPF)
Doctoral research on multimodal generative models and agentic AI for neurological diagnostics, based at CERN in Geneva.
PhD in AI for Neuroscience, EPFL · EDNE Doctoral Program
Initial doctoral enrollment in Lausanne. Research on multimodal Graph Neural Networks, federated learning, and privacy-preserving medical AI for neurological conditions.
MSc in Data Science, Sapienza · 110/110 cum laude
Honors program graduate specializing in Deep Learning and Temporal Graph Neural Networks.
Erasmus MSc, Universitat Politècnica de Catalunya (UPC)
Exchange in Barcelona, focused on machine learning, image processing, and information theory.
BSc in Statistics, Sapienza University of Rome
Statistical analysis, data modeling, and quantitative research methods.
Summer School, ESSAI · University of Ljubljana
1st European Summer School on Artificial Intelligence & 20th Advanced Course on AI (ACAI).
Open Source
Tools I build in the open
Free, AGPL-licensed projects. No telemetry, no accounts, just software that works.
MyTripPlanner
AI road trip planner with the Ulisse agent
Self-hosted planner where Ulisse interviews you and builds itineraries live: stops, map, timings, budget. Runs on your Claude or ChatGPT subscription. No API keys.
TypeScriptNode.jsAI AgentAGPL
MyMacCleaner
macOS system utility with Liquid Glass UI
Open-source macOS app for smart scan, disk cleaning, space lens, duplicate finder, and performance monitoring. Code-signed and notarized.
SwiftSwiftUImacOSAGPL-3.0