Quality anomaly triage
the engagement
client data — not shown
heard
"My inspectors look at everything, which means they look properly at nothing."
real problem
Line sensor data was stored but never modelled. Inspection was a fixed sampling plan, unchanged in eleven years, that took no account of which units were actually at risk.
system built
A time-series feature store over line telemetry, an anomaly model per station, and a triage agent that assembles the inspection queue each shift with the reason for each unit. Inspectors can reject a reason, which becomes a label.
what broke
Two stations had clocks four minutes apart, which the model read as a real signal. Clock drift is now a monitored data-quality check that blocks scoring.
the open rebuild
same architecture · public data
dataset
SECOM semiconductor manufacturing (590 sensors, pass/fail)
what the rebuild covers
Feature store, anomaly model and the queue-assembly logic. The inspector feedback loop is described in the write-up but not built.
artifacts
NotebookCodeDatasetWrite-upDemo