Defect Detection: Under 90% to Over 98%. Six-Figure Annual Savings.
A mid-market industrial parts manufacturer in Detroit, MI with 200 employees and nearly $40 million in annual revenue.
A mid-market industrial parts manufacturer in Detroit, MI with 200 employees and nearly $40 million in annual revenue.
What They Were Facing
A mid-market industrial parts manufacturer in Detroit, MI with 200 employees and nearly $40 million in annual revenue had a quality problem that was eroding their margins and their reputation. Their defect detection rate sat under 90%, which meant roughly 1 in 8 defective parts made it past inspection and into the hands of customers. The resulting returns, rework, and warranty claims cost the company six figures annually in returns and rework, and that figure didn't account for the harder-to-measure cost of damaged customer relationships. The inspection process relied on a combination of visual checks by experienced QA technicians and spot-check measurements. It worked reasonably well when the company was smaller and produced fewer part variants. But as their product catalog expanded to over 300 part numbers and production volume grew, the inspection team couldn't keep up. Fatigue-related errors spiked during the last two hours of each shift, and the company had data to prove it. An equally pressing problem was the tribal knowledge risk. Three of the company's most experienced machinists, who between them knew the quirks of every machine, every tooling setup, and every material behavior, were all within five years of retirement. Their knowledge existed in their heads and in handwritten notes taped to machine enclosures. When a less experienced operator encountered an unusual setup or a material that behaved differently than expected, they'd walk over and ask one of the veterans. The company had no system for capturing, organizing, or transferring this institutional knowledge. Production scheduling was the third challenge. The planning team was using spreadsheets and tribal knowledge to schedule jobs, frequently resulting in machine changeovers that could have been avoided with better sequencing. Estimated changeover waste was running at over 10% of total production time.
Defect detection rate under 90% with roughly 1 in 8 defective parts reaching customers, costing six figures annually
Fatigue-related inspection errors spiking during last two hours of each shift
Three most experienced machinists within five years of retirement with undocumented tribal knowledge
Product catalog expanded to 300+ part numbers overwhelming the manual inspection process
Production scheduling via spreadsheets causing over 10% changeover waste in total production time
How We Solved It
We tackled this as three connected problems, starting with the one that had the most immediate financial impact: defect detection. For quality inspection, we deployed a computer vision system at two critical inspection points on the production line. Cameras capture high-resolution images of each part, and the AI model compares them against trained specifications for that part number. The system flags dimensional deviations, surface defects, tooling marks, and material inconsistencies that fall outside tolerance. Flagged parts are diverted for human review rather than rejected automatically, keeping the experienced QA team in the loop while eliminating the majority of missed defects. Training the model required close collaboration with the QA team. We spent two weeks collecting and labeling images of known-good and known-defective parts across the 40 highest-volume part numbers (covering roughly three-quarters of production volume). The QA lead's expertise was critical here. He identified subtle defect patterns that even the data labels didn't fully capture, and we incorporated his feedback into the model iteratively. For knowledge capture, we built a structured system where experienced machinists document their setup procedures, material-specific tips, and troubleshooting approaches through guided interviews and voice-to-text recording. The system organizes this knowledge by machine, part number, material type, and operation, then makes it searchable through a tablet interface on the shop floor. When an operator is setting up a job on Machine 7 for a specific part, they can pull up everything the veterans have documented about that exact combination. Production scheduling was addressed through an optimization module that takes the weekly job list, considers machine capabilities, current tooling setups, material availability, and delivery deadlines, then generates a sequence that minimizes changeovers. The planning team reviews and adjusts the suggested schedule rather than building from scratch.
Computer vision system at two critical inspection points capturing high-resolution images per part
AI model trained on 40 highest-volume part numbers covering roughly three-quarters of production volume
Guided voice interviews with veteran machinists documenting setup procedures and troubleshooting
Searchable tablet interface on shop floor organized by machine, part number, material, and operation
Scheduling optimization module minimizing changeovers based on machine capabilities and job requirements
Measurable Outcomes
Quantifiable improvements delivered within the project timeline
Improved from under 90% to over 98% defect detection rate
Six-figure reduction in returns, warranty claims, and rework labor
Over a thousand tribal knowledge procedures documented and indexed
Reduced from over 10% to under 8% of total production time
Parts flagged incorrectly, reviewed and passed by QA in under 30 seconds
Fatigue-related defect escape rate reduced by over 75%
The six-figure savings figure represents the reduction in customer returns, warranty processing costs, and rework labor compared to the trailing 12-month period. The remaining gap from the previous cost reflects defective parts that the system catches internally, some of which can be reworked rather than scrapped. The knowledge capture system proved its value faster than anyone expected. Two months after deployment, one of the veteran machinists took an extended medical leave. In the previous era, this would have caused production disruptions on the specialized jobs he typically handled. Instead, less experienced operators used the knowledge base to complete those jobs with a defect rate within a few percentage points of the veteran's typical performance. The plant manager called it the moment the project paid for itself.
Implementation Timeline
A structured approach from discovery to deployment
Camera placement, QA team collaboration on defect classification
Weeks 1-3Camera placement, QA team collaboration on defect classification
Trained on 40 highest-volume part numbers with QA team feedback
Weeks 4-6Trained on 40 highest-volume part numbers with QA team feedback
Guided interview framework and initial veteran documentation sessions
Weeks 7-8Guided interview framework and initial veteran documentation sessions
Changeover minimization module with job sequencing
Weeks 9-10Changeover minimization module with job sequencing
System validation and shop floor staff training
Weeks 11-12System validation and shop floor staff training
Line 1 first, then Lines 2-3
Week 13Line 1 first, then Lines 2-3
System-wide rollout with ongoing optimization
Weeks 14-16System-wide rollout with ongoing optimization
Frequently Asked Questions
How accurate is the AI inspection system compared to human inspectors?
The system catches over 98% of defective parts compared to the previous sub-90% rate with human-only inspection. The tradeoff is a roughly 3% false positive rate, meaning some good parts get flagged for human review unnecessarily. In practice, QA technicians resolve flagged items quickly (under 30 seconds per part), so the net impact on throughput is minimal. The system is particularly strong at catching defects that occur late in shifts when human fatigue is highest.
Can the system inspect all 300+ part numbers?
Currently, the vision model is trained on the 40 highest-volume part numbers, which represent roughly three-quarters of production. The remaining parts are still inspected manually. We're expanding the model's coverage in phases. Each new part number requires a training cycle with labeled images, which takes about two days per part. The goal is full catalog coverage within 12 months of initial deployment.
How do you get experienced machinists to actually document their knowledge?
This was the biggest project risk, and we addressed it head-on. Rather than asking machinists to write documentation (which they universally hate), we used guided voice interviews. A facilitator asks structured questions while the machinist demonstrates the procedure, and the audio is transcribed, organized, and formatted automatically. Each session takes 15-20 minutes and covers one procedure. The veterans were skeptical at first, but once they saw their knowledge organized and accessible to younger operators, they became advocates. Several started volunteering procedures we hadn't asked about yet.
Does the scheduling optimization account for rush orders and unplanned downtime?
Yes. The system re-optimizes when priorities change, which happens frequently in a job shop environment. Rush orders can be inserted with a priority flag, and the system recalculates the sequence to accommodate them with minimum disruption. Unplanned machine downtime triggers automatic rescheduling of affected jobs to available machines with compatible tooling. The planning team retains override authority on all scheduling decisions.
Services Used in This Project
Ready for Similar Results?
Schedule a discovery call to discuss your specific challenges and learn how we can deliver measurable outcomes for your organization.
Schedule Discovery Call