AI as part of the system, not a separate one.

Every model is checked when it is uploaded: does it load, does it run, and — optionally — does it give the same results as in training. A released model cannot be changed afterwards, and the check result is stored and can be repeated.

Each model also states which input values it expects, in which order. The setup assistant checks this against the live data before you can continue — so a model never runs on the wrong inputs.

1. Name the evaluation.
2. Choose the inputs — controller values, sensors, calculated values.
3. Decide when the model looks (on by default): only data from the actual process reaches the model, and if that information is missing, the model gets nothing.
4. Choose what is calculated from each input.
5. Select the model; the assistant checks that it fits.
6. Define the outputs — score, class, load estimate.
7. Review the result, then assign it to machines and choose where it runs.

Models with several sensors process each sensor right where its data is, and only small results travel. If any sensor is silent, there is no result at all.

A model runs on the EdgeBox, on an AI box or centrally. Toleron chooses based on where the data is and what each device can do, follows your preference where possible and records why whenever it has to deviate. An EdgeBox is only allowed to run models once that has been verified on the box.

Every new model first runs in a test phase in the background and can only go live after an observation period. Monitoring that would stop a model and raise an alarm if its results gradually shift — and would never quietly re-adjust it — is designed, but not running today.

Developed offline on real data from our lab's tool grinder: a sharpness index for the wheel, calculated from sensor data during contact, and a decision layer that turns it, together with the operator's own measurements, into advice on when to dress. The model runs on the AI box, and its advice appears in Toleron like any other value.

Upload manuals, work instructions and specifications per machine. Then ask in plain language — on the machine's page or in the side panel that is always available. Answers state how confident they are and name the source documents. Everything runs on your own hardware, so factory documents never leave the factory, and the assistant declines rather than guesses when there is no data.

Toleron Flow answers questions about your operating data — alarms, downtime, recordings, standards — and is in early access on our lab installation. It names every source and never interferes with data collection.