A model that looks fine on a sample set and misses the defect you ship is a failed release. We develop it on the line's failure cases, then build the application into the quality work people already do.
Six parts. Five should be caught. Click one, then swap the model. The count is failures caught. The note is why the call moved.
The model is the small part. The result is a part held, a person who signed, and a line that can keep running when the model is unsure.
Computer vision and related models for the defect, the part, or the condition you care about. Scored against the failures from your line. Replaced when a better model wins on those cases. The panel above is that swap on six example parts.
The application around the inspection. What gets held, who releases it, what goes back to the line. Quality still owns the decision when the model is short of the bar you set.
Screens and agents on the systems already in the operation, including on premises when the images and the process data stay at the plant. The people who develop the model are the people who test it.
We start with the parts that should have been caught. Those cases decide the release. A miss after go-live is kept and added to the set the next model has to beat.
If we cannot measure it on the line's actual failures, it is not ready. A clean score on someone else's images does not count.
The model can flag. It does not ship a bad part on its own when the call is past the authority you set. Someone on the line or in quality stops it.
The application is left in a state quality and the plant can operate. We do not stay in the loop as a condition of the system working.
Images, process data, and model artifacts stay yours. They are not training material for a product we sell to the next plant.
Scott Ancheta spent two decades at Cisco and Dell deploying systems for manufacturers including GM, Ford, and Fiat Chrysler. He was on site when they went live. SRA's AI work is held to that same expectation. The application has to run in the plant, and the people who build it are the ones who test it there.
Where the application also has to satisfy an auditor, the regulated release is the same standard: policy, a named person, and a record.
Yes. The model is scored on the line's actual failure cases. A sample set from somewhere else does not decide the release. If another model catches more of your defects, we switch.
Yes. The application is built for the systems already on the floor, including cases where the data and the model have to stay on premises.
A person can stop the action. The miss is kept, because the next version is measured against it. Quality still owns the release.
Regulated industry, finance, and e-commerce.
We'll tell you what we'd develop on those cases, how the application would sit on the floor, and what has to be governed before anyone relies on it.