Senior data scientist and ML engineer at Inmatica. I build open-source tools for clinical and scientific work, mostly local first.
My route here went through the lab: a bachelor's in biotechnology at the University of Genoa, then a master's in advanced biotechnology and a PhD in engineering sciences at the Vrije Universiteit Brussel, where I worked on CheckPack, building machine-learning models for sensor-based food-spoilage detection. Most of what I build now comes out of that experience. Scientific data is messy, and a tool that hides how it reached a number is hard to trust.
Most of the applications below share a shape: a local Python service with an Angular or React interface, and a model that either ships with the app or points at an endpoint you already have.
DILIGENT Clinical Copilot Python FastAPI Angular
Decision support for Drug-Induced Liver Injury review, running on one machine. You enter a case, run an assessment with a local or hosted model, then check the draft against the structured findings, medication timeline and drug-match flags before anyone acts on it.
XREPORT Radiological Reports PyTorch Transformers FastAPI Angular
Generates editable draft radiological reports from X-ray images, and covers the loop around that: dataset preparation, local training and evaluation, side-by-side comparison of several report generators. A research tool, not a diagnostic device.
EMADB Autopilot Python ChromeDriver
Collects adverse drug reaction reports from the EMA's EudraVigilance database by driving the search interface and saving the exports. No longer maintained.
ADSMOD Adsorption Modeling Python PyTorch React
Adsorption workbench. Fits theoretical isotherm models to experimental data, enriches the material records with PubChem properties, then trains predictors on the NIST and ARPA-E isotherm collections.
LLMeter Rust
Single binary CLI for measuring local OpenAI-compatible providers: repeatable benchmark runs, performance scenarios, and Markdown or HTML reports you can hand to someone else.
TKBEN Python FastAPI Angular
Compares tokenizers on your own text. Vocabulary, throughput, and whether encode/decode actually preserves the input, with the dashboard exportable as a PDF.
ParaGraph Python FastAPI React Flow
Visual workspace for LLM workflows. Wire the steps together as nodes, run the graph, and watch each stage as it executes, which is easier to reason about than a chain of prompts buried in a script.
AEGIS Geospatial View Python FastAPI Angular
Geospatial workspace that starts from a question instead of a form. Ask for a place, a set of coordinates or a mapping task, and it assembles the view from the public data sources it has access to.
FAIRS Roulette Player PyTorch FastAPI React
Workspace for reinforcement-learning experiments on roulette. Import or generate data, train a DQN agent, save checkpoints, then step through inference sessions and compare each suggestion with what the wheel actually did.
FeXT AutoEncoder PyTorch PySide6
Desktop pipeline for training convolutional autoencoders over image datasets and exporting the latent vectors, built on a VGG16-style encoder. No longer maintained.
Python and Rust for most of the work, TypeScript for the interfaces. PyTorch, Hugging Face Transformers and scikit-learn for the models; FastAPI, SQLAlchemy, SQLite or PostgreSQL, Docker and GitHub Actions around them. On the science side: RDKit, NumPy, pandas, and public databases such as NIST, ARPA-E and PubChem.
