Blog
Field notes from the platform
How we build, simulate, and operate heavy industry — advisory AI, server-driven control screens, and the architecture decisions behind them. Written for the people who run plants, not the people who buy slideware.
QME 2026, wrapped: real water, local inference, three days of hard questions
Our stand at the Queensland Mining & Engineering Exhibition ran a live water filtration rig beside the closed-loop simulator — with the advisory AI doing its inference on hardware at the stand, not in our cloud. What we showed, what it previews, and what the floor asked us.
Where the standard library goes next: water, gas, and data centres
The shipped library is 22 typed device types across three libraries. Next: full control device libraries for water utilities, gas, and data centres — planned, sequenced, and shaped by what beta users ask for first.
Bad data must never look live
A value can be wrong two ways: old, or broken. Screens that collapse both into a last-known-good number train operators to trust a lie. Every value on our wire now carries one of five quality states — and "bad" can never be rendered as live, stored as history, or satisfy a command precondition.
Dead chain, idle plant, or frozen PLC? Teaching the platform to tell
Every SCADA screen has shown green over a dead link at least once. A newest-sample-age check cannot tell a quiet plant from a broken collection chain from a controller that stopped scanning — so every controller now carries a heartbeat that can. It is the newest type in the standard library.
C#, not Structured Text: retargeting the PLCnext generator
We built a deterministic generator that emits IEC 61131-3 Structured Text for PLCnext — then retired that target. The controller runtime told us something better was available: the same C# blocks the simulator already runs. Here is the finding, and why one behavioural source now feeds every target.
Sign in when the internet will not: identity that survives the link dropping
A mine site cannot wait on a cloud round-trip to let an operator log in and acknowledge an alarm. So identity runs three schemes behind one selector — cloud single sign-on when connected, site-issued PIN tokens when the link is gone, and keys for machines. Here is how we drew that line.
The plant without the plant: a simulation that runs the real control logic
A bad control test does not roll back — it surges a duty/standby changeover or trips a VFD on equipment people are standing next to, and the commissioning window is short and shared with production. So the platform ships a native, licence-free simulation that runs the same control logic as the PLC, addressed exactly like one. Here is how we drew that line.
Generated, not written: deterministic PLC code from the device model
Hand-written PLC code drifts from the HMI, the tag database and the docs — the same pump described in five places, each one diverging, then re-typed at the next site. We generate the control code from the same device model that draws the faceplate and feeds the historian, deterministically, so it cannot drift. Here is how.
One device model, many PLC vendors
Every site is a different mix of PLC brands and vintages, and the usual fix is a vendor stack per plant. We speak to a tag, not a vendor — one device model over a thin, swappable driver seam, with a hand-rolled EtherNet/IP CIP stack so the edge leans on no vendor SDK. Here is why we drew the line there.
Edge-first: why a site keeps running when the link to the cloud drops
A Consystence site runs at the edge and treats the cloud as a vantage point, not a dependency — so the plant keeps working when the link drops. Here is how we drew that line, and where the field record still has to catch up.
One semantic UI tree, many renderers
Our control screens are not an SPA. The server drives every screen, and holds one semantic UI tree the CLI already renders live — with the operator faceplates and a native renderer following. Here is how that works and why we chose it.
Advisory by design: why our AI explains the plant but never controls it
AI on a Consystence site reasons about the plant and explains anomalies — but it cannot command equipment. Here is why that line is drawn in the architecture, not the marketing.