On-premises AI
On-premises AI that learns from your corrections
The engine that reads your rack photos runs on your premises. It improves with your teams’ decisions, and a person always has the final say.

Contents
An engine that runs on your premises
The analysis engine is installed on your servers, along with the rest of the platform. It calls no online service: photos are analyzed on site and never leave your network.
What it reads on a photo
- The family of each piece of equipment: network, fiber and transmission, power, server, patching…
- Its position in the rack, in U.
- A confidence level, shown next to the result.
Uncertain results are flagged
When the engine is not sure, the result is marked “to check” with the reason. It is never presented as certain.
Learning from your corrections
Every discrepancy your teams approve or correct is kept: the photo, the bounding box and the confirmed family join the training set. The next model thus learns from your equipment, your racks and your rooms. Training runs on a graphics card, on your premises or by our team.
A new model only if it performs better
Before it is activated, a new model reads the same control photos as the model in service, and the results are compared. It goes into service only if it performs better; otherwise, the model in service stays in place. Nothing degrades silently.
A person always decides
The engine proposes, a person decides: nothing is written to the inventory without approval, at one or two levels. See also our page on the EU AI Act.