What a Google Cloud hackathon let us build
The moment it could all have stopped
In the middle of the work, our credits on a paid large model dropped to zero. Not figuratively: the balance was empty, and without a large model, an agent project slows to a crawl. That is when the hackathon mattered: the credits offered by Google Cloud gave us back access to a large model — Gemini — at the exact moment we had none left. Gemini held the post when nothing else could.
What we built with it
Not a throwaway demo: a system that still runs.
- A control tower — a cockpit to make intelligent agents work together, with guardrails: gates that can refuse an action, circuits that demand proof before validating.
- Alice, an agent that learns — small, local, made of simple organs: a router inspired by network routers, an external memory, eyes that read images, a heart that adapts its tone. She keeps what she learns outside the model's weights, and it can be verified.
- A method — every capability laid down with a test written first, a replayable proof afterwards, and the honesty to write down what is not yet proven.
What the back-and-forth taught us
Switching from one model to another — Gemini, then others, depending on what we had — was not comfortable. But it is precisely that back-and-forth that made us understand the inner workings of our own system, and fix what did not hold. We could work at home, on modest machines, and run our studies with what we had at hand. It was not easy. It was thrilling, exhausting, addictive. And a great deal of gratitude came out of it.
Thank you, and what comes next
Thank you to Google Cloud for the contest, the credits, and the model. This helping hand did not just unblock a project: it shaped a way of working — sovereign, local, verifiable — that we keep digging into. The door opened at the right moment. We fully intend to keep walking through it.