Housekeeping compliance was tracked through manual photo inspection across 22 cities. Operations managers could wait 12 hours before learning that an inspection had failed, often after the resident experience had already suffered.
The obvious automation was not the whole problem.
Increasing automation alone could create false passes at scale. The system had to improve coverage and speed while keeping inference cost controlled and ambiguous evidence visible to human reviewers.
The important threshold was operational, not academic: at what confidence could the system act, and below what point should a person make the call?
Build the operating loop, not only the intelligent step.
- 01
Translated the housekeeping checklist into a repeatable visual evaluation flow.
- 02
Used CPU inference to keep the operating cost within the deployment constraint.
- 03
Established a 65% confidence gate below which evidence routes to human QA.
- 04
Added drift alerts after a real-world distribution shift exposed silent accuracy degradation.
Audit coverage moved from periodic spot checks to a full-cluster daily loop. Operators gained a much faster view of failures without allowing low-confidence model output to pass as certainty.
What stayed after shipping.
Model quality is not static once it meets the world. Lighting, device behaviour, and operating changes are product inputs; drift monitoring is part of the user experience.