Case histories
Not concepts. Shipping hardware, live platforms and models in daily use — three of them built by this team from the enclosure up.
Case 01 · alyt.com
A voice assistant, a security system and an open hub in one quiet object. Speech, reasoning and replies run on hardware inside the house. Hosted by us, or entirely by the owner. Luca founded ALYT and runs it today.
Every layer in one programme: the enclosure and its light guide, the board and its radios, the local speech and intent models, the firmware, the mobile and TV apps, and the multi-tenant platform underneath — with a developer surface so other people can build on it.
It is the reference implementation of everything this studio argues for. The house answers in the house; the cloud keeps the account and delivers the next model.
Case 02 · gpsx.ai
GPS Standard has been making perimeter detection in the Aosta Valley since 1974 — the company name comes from its Ground Perimeter System patent, buried detection that sees without being seen. GPSX is its edge-AI generation: the same physics, with the decision moved into the sensor.
Perimeter systems are judged on one number that never appears in the brochure: how often the operator believes the alarm. Local classification is what moves that number, because the discrimination happens with the full signal, at the fence, in real time.
GPSX runs on the same platform as ALYT — tenancy, provisioning, telemetry and connectors inherited rather than rebuilt. That is the argument for owning your platform: the second product costs a fraction of the first.
Also in the building
AROMI.AI
A connected coffee platform: ESP32 firmware, an in-house PCB redesign onto a single board, a mobile app, a web dashboard and a cloud backend — plus a connector so it behaves like a citizen of the ALYT home rather than another lonely app on a phone.
FINANLY.AI
Finance intelligence and close automation, where the interesting constraint is not model size but confidentiality. Local models doing structured financial reasoning over data that was never going to be allowed out of the network.
THE MODEL FACTORY
Data generation, training, evaluation and release, built as repeatable machinery rather than a folder of notebooks — so a new domain model is a scheduled piece of work instead of a research adventure. It is the reason we can promise a custom model with a straight face.
Next one
Everything on this page started as a conversation about something that had to happen faster than a network round trip. Yours can too.