Turning a node amber before anything is wrong with it
A mining value chain runs from pit to port. A delay in one link lands as a loss three links downstream, half an hour later. I designed the interface that shows the damage before it arrives.
Every project turns on one question: is the obvious explanation the right one?
Teams cannot see enough. Bring more data into the operations centre, put it on bigger screens, and supervisors will catch problems earlier.
They had plenty of screens. The failure was that no deviation looks like a problem where it starts. Trucks queuing at the pit is a haulage number. The loss lands at the crusher thirty minutes later. Nobody owns the space between two links, so nobody was watching it.
I mapped what actually happens when an alarm fires. Detection was never the slow part. Assessment was. A supervisor would spend twenty to thirty minutes on the phone reconstructing a root cause from four different systems, by which point the shift target was already gone. That reframed the product. It was not a monitoring problem. It was a problem of connecting cause in one place to consequence in another, fast enough to still act.
From 1.8 min to 2.4 in the past 15 mins.
Feed rate is 0.3 tph below target
This stockpile is at risk due to downstream effects of the Haulage queuing at the crusher.
Rate is below target due to Crusher Hauling bottleneck
Colour the future, not the present
When haulage goes red, the crusher turns amber. Nothing is wrong with the crusher yet. It is amber because the model says it will be in thirty minutes.This is the whole point of a prediction. If the interface only paints current readings, the forecast may as well not exist. Amber has to mean coming, not happening.
Start everything grey
The default state of the value chain is neutral. No status colours, no green, no decorative accents anywhere in the chrome.If the resting state is colourful, colour stops carrying information. Reserving it entirely for deviation is what makes a single amber node readable across a room.
Give the AI one fixed shape
Every generated message follows the same four parts, in the same order, in the same place: what the issue is, why it is happening, what it will cost, what to do about it.Under time pressure people read by position, not by reading. A fixed structure lets a supervisor jump straight to the action. It also forces the model to show its reasoning rather than assert a conclusion.
Put the camera next to the number
A contextual video stream appears alongside the alert, showing the actual queue at the pit or the actual feed level at the crusher.Operators trust what they can see. The video is the fastest way to confirm the model is describing something real.
Treat the handover as an output
Issues, decisions and outcomes accumulate into the shift handover automatically rather than being retyped.Knowledge was living in individuals and dying at the end of each shift. If the system does not remember what happened, the same deviation costs the same money next month.
Haulage bottleneck that could impact Mill 2 feed rate. Would you like to review the current forecast and proposed mitigation?
You05:02PMShow me the drivers of the forecast drop.
Crusher feed rate has dropped below 7,000 TPH because of increased queuing at the dump area. Haul cycle has increased from 20 to 23 minutes in the past 30 minutes
AI can provide incorrect information. It is important for all users to verify all information before use.
All clearore moving, nothing flagged
Redrawn from the real screens. Grey is normal. Red is a live deviation. Amber is a link that is still healthy but will not be shortly.
The argument redrawn: a haulage delay flags red at the pit, and the crusher goes amber thirty minutes before anything is actually wrong with it.
Everything starts greycolour is reserved for deviation
Key Alerts & Follow-Ups
Scheduled maintenance inspection for 08:00. Continue monitoring trend; no immediate shutdown.
Flagged for calibration check during next maintenance window. Operator instructed to manually log fuel metrics.
Auto-recovery executed via Nucleus logic. System confirmed stable. No further action required.
Auto-corrected by blending optimizer. System recalibrated post-shift; no spec impact.
Incident under investigation. Safety team to review footage and conduct crew debriefing. Temporary access restrictions in place.
A prediction shown as a number is a fact. A prediction shown as a state change on the thing it will affect is a decision.
Most AI product work stalls in exactly that gap. The model can predict something useful and the interface has no way to say it that a person can act on. That is design work, not model work, and it is usually where the adoption problem actually lives.
Have a problem like this one?
Send a link and a sentence. If I can help, I will record a teardown. If I cannot, I will say so.
Start a conversation