InternalAutomation

Case Study · Manufacturing

Vision-based defect detection for a manufacturer

A custom computer-vision inspection system caught defects manual checks missed, cut escapes to customers, and freed inspectors for high-judgment work.

Client
A precision parts manufacturer, 3 lines
Market
Remote, US
Timeline
11 weeks to launch

Anonymized and illustrative of a typical engagement.

−72%
defect escapes to customers
3x
inspection throughput
100%
of units inspected, not sampled
11 wks
to production

01 / The challenge

Where the time was going

  1. 01Quality relied on human spot-checks and sampling. Inspectors were skilled but could not examine every unit at line speed, so subtle defects slipped through and reached customers as costly returns and reputation damage.
  2. 02Sampling meant entire batches could ship with a recurring defect before anyone noticed the pattern.
  3. 03The best inspectors were spending their time on repetitive visual checks rather than root-cause work.

02 / The build

What we shipped

We built a custom vision system that inspects every unit at line speed and routes uncertain cases to people.

  1. 01Custom defect modelA vision model trained on the manufacturer's own parts and defect types, not a generic detector.
  2. 02Edge inferenceInference runs on the line at speed so every unit is inspected without slowing production.
  3. 03Human-in-the-loopBorderline cases are routed to an inspector, whose decisions feed back into the model.
  4. 04Pattern alerts and MES integrationRecurring defects trigger alerts and feed the manufacturing execution system so root causes get caught fast.

03 / The results

What changed

Quality went from sampled to total.

Defect escapes to customers fell 72% because every unit is now inspected rather than a sample, and inspection throughput tripled. Recurring defects are caught as patterns early, and inspectors moved from repetitive checking to root-cause analysis.

−72%
defect escapes
100%
units inspected
3x
inspection throughput

We went from checking samples to checking everything, and our inspectors finally work on why defects happen.

Quality Manager, manufacturer

05 / FAQs

Questions about this build

Will it slow the line?

No. Inference runs at the edge at line speed, so every unit is inspected without reducing throughput.

What about new or rare defects?

Borderline and novel cases route to an inspector, and their decisions retrain the model so coverage improves over time.

Want a result like this for your team?

Name the work that is costing you the most time. We will map the build, show what is worth doing first, and what it costs. If there is no fit, we will say so.