Augur
An agentic predictive-maintenance system for the plant that keeps a data centre running.
A tool for deciding when to service a machine, rather than waiting for it to break.
Every pump, chiller and switchboard fails eventually. Service everything early and you throw away parts with years left in them; wait for the failure and you pay for the breakage and all the downtime around it. Most operations cannot tell which of the two they are doing.
Augur ages a set of equipment along its real failure curve, lets you choose a maintenance strategy and how much sensor coverage to pay for, and prices the outages you avoided against what the monitoring cost.
of impactful data-centre outages start in the power train. Most often the UPS.
Uptime Institute, Annual Outage Analysis 2025.
of operators say their most recent significant outage cost more than a million dollars. More than half say it cost over a hundred thousand.
Uptime Institute, Annual Outage Analysis 2025.
Critical plant is
maintained to a calendar.
The calendar decides when a machine needs attention months before the machine does. Sometimes that is early and a good asset is stripped for nothing. Sometimes it is a week late.
The window
Machines give warning before they stop. The warning is physical and measurable, and it arrives long before anything trips.
Between the first detectable sign and the moment the machine stops, there is time.
Reliability engineers have called this the P–F interval for fifty years. Its length is a property of the failure mode: bearing wear gives weeks, an arc fault gives seconds.
Every hour of that window you can see is an hour you can act in.
Augur finds the window early, works out what it is worth, and spends it.
The system
Four layers over plant that is already instrumented for most of it. Most of the sensors are on the floor. Little reads them continuously.
Four layers, one register, an agent for every subsystem.
Instrument
Vibration, motor current, temperature, battery impedance and differential pressure, read continuously off the BMS and the sensors already fitted to the plant.
Model
Every asset carries its failure modes, its P–F intervals, its duty history and what an hour of its downtime costs, in one register the agents read from.
Judge
An agent per subsystem watches its own signals, checks them against the neighbouring plant, and estimates how long the asset has against the failure mode it thinks it has found.
Act
It raises the work order, reserves the part, books the window against the load profile, and tells the engineer what it saw and how sure it is.
Each subsystem gives a different amount of warning.
Lead times are published predictive-maintenance benchmark ranges for this class of plant. They are what the design assumes, not measurements from one site.
How it is run
You decide how much of it runs without you. That is a setting, and it moves as trust does.
Three ways to run it.
The system is identical in all three. What changes is how much of its own decision it carries out before a person sees it.
In the loop
It proposes. An engineer approves every action before anything moves. Useful while the floor is deciding whether it believes the model.
On the loop
It acts inside limits you set: routine work orders, parts, scheduling. Anything touching live plant or spend above a threshold comes back to a person.
Out of the loop
It runs the routine work and calls you for the exceptions. The engineer's day becomes the hard calls and the escalations.
A ranked list, and the reason for the ranking.
Ranked on criticality multiplied by imminence, so the work that protects the most load and has the least time left sits at the top. Illustrative readings.
What changes
Faults are found while the machine is still running. The work moves into planned windows, and the part is on site before the engineer is.
The benchmark ranges the model is anchored to.
Published ranges for condition-based programmes on this class of plant. Your own figure depends on your outage cost per hour, your sensor coverage and how much of the estate is already past its design life.
You can run this one yourself.
Set the maintenance strategy, choose how much sensor coverage to pay for, then watch five subsystems age along the curve. It prices the downtime you avoided against running the plant until it breaks.
Open Augur →A studio for documents that have to work the first time they are read.
hello@enzwa.com · enzwa.com · Hong Kong