Hi, I'm Valerio.
I'm an AI engineer and tech lead in Zurich. I build agent harness systems: the orchestration, the tools and the evaluation that an agent needs to work every time, not only in a demo. I also fine-tune models for specific domains. I use supervised fine-tuning or reinforcement learning, depending on what the task needs. And I work on foundation models for structured data. Before AI, I studied physics and did a PhD in electrical engineering at KAUST.
Working on an AI project? Agents, fine-tuning, evaluation or AI security: email me with a few lines about it.
Selected work
-
A generative model for relational health records
A model that learns from hospital records as linked tables, ordered in time, and writes synthetic patient records with the same structure.
-
Reinforcement learning for an autonomous pentesting agent
I fine-tuned Qwen 32B with reinforcement learning for a multi-agent penetration-testing system. The training worked, and the frontier models still moved faster. Here is when fine-tuning pays.
-
AI coaching from poker hand histories
Tech lead on a coaching product. I contribute to its agent architecture and have shipped several of its AI coaching features. A system of several agents turns raw hand data into short, specific feedback that a player can act on.
-
SDAC: assessing synthetic health data
SEARCH is a European project for health research with synthetic data. I work on the part about trust: does a synthetic dataset protect the people behind the real data, and is it still good enough to use?
Personal projects
Things I build in my own time. Usually I want to know whether something is possible before I have an opinion about it.
-
autome
A personal AI assistant in Telegram. Send it messages, voice notes, photos and documents. It files them into a wiki that you can ask questions of, and keeps trackers, reminders and shared spaces.
-
The Primer
An AI tutor that teaches young children reading, writing and arithmetic. The curriculum lives in an engine outside the model. The model leads the conversation and does not decide what to teach.
I also built MISTERO, an AI murder-mystery game, and Bussola, a platform for neutral analysis of public debates.
Writing
Small experiments that take a popular claim about AI and test it. You can run each one yourself.
-
Can a decision model find where an agent failed?
I tested Jev, a decision model, on 257 failed agent runs that people labelled by hand. On runs that fit in its window, it finds the failing step about as often as a frontier model. It cannot name the kind of failure. It knows when it is unsure.
-
Confident and wrong
When can you trust a model that says how sure it is? I built a toy security world where the right confidence is known. Two of three training methods became honest, but only with enough data. Unfamiliar input broke all three.
-
Is visual data really richer than text?
A picture and a list of symbols that carry exactly the same information about a toy world. The picture model learns better. Take away its convolution and the advantage is gone.
-
Nobody told it what a noun is
A very small language model reads four novels for five minutes. Nouns, verbs, adjectives and adverbs separate inside it, and nothing in the training mentions grammar. With shuffled word order, they do not.