Inference at the wellhead
Models sized for the hardware that is actually on site.
NSK·SOL·06
The site has a gateway with eight gigabytes of memory, intermittent backhaul, and a standing rule against sending video off the plant.
The useful model was trained for none of that. What matters is how much precision can be given up before it stops being right.
Set the precision budget.
Set the precision budget. Watch which digits survive it.
Read at the gateway, with counts and alarms crossing the boundary and the video staying on the plant.
Work orders, manuals and shift logs searched together, so a fault is met with what was done last time.
The log written from the shift's own records, for an operator to correct and sign.
Site safety cases updated for advisory-only output. No model sits in the control loop.
Footage stays on the plant; only counts and alarms cross the boundary.
The alarm remains the operator’s decision, and is recorded as such.
Bring us the case you have not been able to close.
We will tell you which part of it is a model problem and which part is not.