I build the identity and infrastructure layer that makes AI agents useful on real code.
Software engineer, twenty years of shipped systems across public administration, enterprise, advertising, games and regulated betting. The breadth is the point: it is what makes the current direction credible rather than fashionable.
Things that exist and run. Links go to the real thing.
A CLI that reads your own LinkedIn data through the EU Digital Markets Act portability API — the route almost nobody documents, because LinkedIn's official API will not give you your own profile. Version pinning, unreliable pagination, encrypted credential storage. The README is the write-up the official docs are missing.
A real pizzeria in Italy, with real customers and a real menu that changes. Customer-facing PWA on one side, a self-hosted CMS the staff actually use on the other, plus an internal shift and checklist app. Nobody's side project: if it breaks, someone's evening service breaks with it.
Five AI coding agents sharing one instruction layer, one memory, one tool registry — with a machine-identity registry that generates every agent's configuration from a single source, secrets encrypted at rest and documented by pointer only, and scheduled integrity checks. Built because running agents seriously is an operations problem long before it is a prompting problem.
A status and management dashboard for the VPS that hosts all of the above — written in Go with templ and htmx, no SPA, no build pipeline worth the name. Server-rendered HTML turns out to be enough for almost everything I actually needed.
A cooperative graphic adventure built for my two sons — five missions and a hub, installable, no accounts and no ads. The design constraint was that two children of different ages have to be able to play it together without one of them losing. Development stopped; it stays online because they still play it.
For the Italian Ministry of Foreign Affairs: online consular services for Italian citizens living abroad. Java EE, JBoss Seam, Oracle. Still in production more than a decade later, which is its own kind of review — and the reason I take long-lived systems seriously.
Occasional, and only when I have something I have actually tested.
Machines are getting very good at answering "how". They cannot answer "why", and I do not think that is a temporary limitation. What that means for what I teach my children.
ComingHow five agents on three machines get identities, what a leaked token actually costs, and why the MCP authorization spec turns out to be plain OAuth 2.1 wearing a new hat.
It started with an Intel 386 and never really stopped. Twenty years later the through-line is still the same one: I like understanding a system all the way down, and I like the part where you get your hands dirty — hardware, electronics, a soldering iron, code that runs on a machine you can hear.
The professional record is deliberately broad — public administration, a global agency, enterprise consulting, video games, and six years in regulated sports betting. I used to think of that as a lack of focus. I have come round to the opposite view: generality is a skill, and it is the thing that makes a new direction credible instead of opportunistic. Somebody who has only ever shipped one kind of software claiming they now do AI is guessing. I have shipped six kinds.
What I care about now is narrower. AI agents are being handed real access to real systems far faster than anyone is working out how they should be identified, authorised and audited. That gap is interesting, it is unglamorous, and it is exactly the kind of problem twenty years of operations makes you good at.
Outside the screen: two sons who get most of the good ideas, a workshop, and a standing preference for things built over things assembled.
Interesting problems, awkward questions and disagreements are all welcome. So is a straight "we are hiring for this, are you interested".