QME closed its doors on 23 July, and we have just about caught our breath. Launching the public beta the day before a trade show and then standing on the floor for three days was an ambitious way to spend launch week — and exactly the right one. Thank you to everyone who stopped at Stand C820, asked an awkward question, or made a pump trip on purpose to see what would happen.
What was on the stand
Two loops, side by side, running the same platform.
The first was real: a small water filtration process built into the stand — actual water, an actual pump, actual instruments — with its equipment authored from the same typed device library that ships in the beta. Live tags, live alarms, live screens assembled in real time, exactly as they would be on site.
The second was simulated: the closed-loop sim runtime running control logic the same way a physical PLC does, the setup we have written about in The plant without the plant. Same device types, same faceplates, same screens — one fed by water, one fed by physics.
Putting them next to each other was the whole argument in two square metres. Visitors could throttle the real rig and watch the platform read it, then walk one step left and do the same thing to a plant that does not exist. If you cannot tell the screens apart, the simulation story stops being a slide and starts being a tool.
The AI never left the stand
The part we were most keen to prove: the advisory AI on the stand ran its inference locally, on hardware sitting under the bench — connected to both the filtration rig and the simulator, with no cloud round-trip in the loop. It read live tags and alarm state, watched visitors mistreat the rig, and explained what the process was doing as it happened.
Partly that was trade-show pragmatism — exhibition-floor connectivity is its own hazard study. But mostly it was a preview we wanted to give in person. A lot of the sites we build for cannot, or will not, backhaul plant data to anyone’s cloud, and our roadmap says the advisory tier should still work there: on-device inference at the edge, for exactly those sites. QME was the first time we demonstrated that pattern live, against a real process, in public.
And one line held on the stand exactly as it holds in the product: the AI explained the process and never controlled it. The filtration sequence, the interlocks, the trips — all of that lived in control logic, the way Advisory by design says it must. The AI’s job was to narrate: why the discharge pressure was climbing, what a trip had actually tripped on, what it would suggest — and a suggestion is where it ends.
Shipped versus demonstrated
We label our claims carefully, so to be precise about last week:
- Shipped today: the platform the stand ran — the typed device library, closed-loop simulation, server-rendered control screens, and the tiered advisory AI. That is the public beta; the sandbox is free.
- Demonstrated at QME: the advisory AI running on local inference hardware against a live process and a simulator. Proven live on a stand — not yet a shipped, supported deployment mode.
- Roadmap: productised on-device inference on edge controllers, alongside the on-device anomaly detection already on the solutions page. The stand was a preview of that future, not a delivery of it.
What the floor asked us
Three questions came up over and over, and they are worth answering in public. “Does the AI write to my plant?” No — advisory in every tier, nothing executes without operator confirmation, and hard interlocks stay in IO and safety relays. “Does this replace my SCADA?” Not today — beta means stand it up alongside, author and simulate, and keep production control where it is until we take the label off. “When do you do water?” Sooner than we expected to be asked — the interest from water and wastewater people was strong enough that it firmed our conviction in the standard-library roadmap.
If you saw the rig and want the same demo pointed at your own process, book a time — the calendar is considerably quieter than the stand was.