Most of the machines that keep a building, a farm or a small factory running have no continuous intelligence attached to them. A motor either works or it does not, and the first signal of trouble is usually the outage. We have spent the last few months building toward a different answer, and this post explains the reasoning.
The gap we keep running into
Monitoring hardware is not the bottleneck. Sensors are cheap and gateways are everywhere. The bottleneck is configuration. Every installation needs someone who knows the machine to define what normal looks like, wire the tags and tune the thresholds. That work costs more than the hardware, which is why the smaller half of the installed base never gets monitored at all.
Three layers, not one model
Our platform is structured in three layers. Signal intelligence collects and conditions current, vibration, temperature and acoustic data. Equipment intelligence learns how that specific machine behaves and scores deviations against its own history. An explanation layer turns structured findings into something a technician can act on.
The separation matters. Anomaly detection is a numerical problem with a deterministic answer. Explanation is a language problem. Collapsing them into one model makes the system harder to test and impossible to certify.
Why the language model never touches the machine
A language model is the wrong tool for a control loop. It is non-deterministic, it is slow relative to a sampling loop measured in microseconds, and its output varies with phrasing. In our architecture the deterministic layer decides, and the language layer describes. The boundary between them carries structured findings only, never raw samples and never commands.
This is not a limitation we plan to engineer away. It is the property that makes the system reviewable.
Where we are today
We are at the bench prototype stage: sensor integration, data capture and equipment-state recognition on a test rig. Baseline learning and field alerting come next. We are being deliberate about not describing target capabilities as validated results.
You can read the full architecture, the concept enclosure and the development roadmap on our Physical AI page.