99.2%
defect-detection accuracy vs. golden set
-70%
manual inspection hours
3x
throughput at the inspection stage
~9 months
to project payback
The Client
A precision parts manufacturer producing machined and stamped metal components for automotive and industrial customers. Several production lines running two shifts, with contractual quality requirements measured in single-digit defective parts per million — the kind of tolerances where one escaped defect can trigger a customer audit.
The Challenge
Every part passed through manual visual inspection: trained inspectors checking for surface scratches, burrs, dimensional anomalies and coating defects under bright light, hour after hour. It was the slowest station on the line and the hardest to staff — inspector fatigue meant accuracy drifted across a shift, and experienced inspectors were retiring faster than replacements could be trained.
As order volumes grew, the choice was stark: add a third inspection shift at significant recurring cost, or automate. An earlier experiment with an off-the-shelf vision system had failed — too many false rejects on acceptable cosmetic variation, and no way to tune it. The client needed a system trained on its own parts and its own defect taxonomy, running at line speed on the factory floor, without shipping images to the cloud.
What We Did
Dillo delivered the system as an outsourced AI project: a team of computer-vision engineers, an ML engineer and a delivery lead, working milestone by milestone with the client's quality manager as product owner. We started not with models but with data — building a labeled image library from the client's own lines.
- Instrumented inspection stations with industrial line cameras and controlled lighting, capturing tens of thousands of part images that inspectors labeled against the client's defect taxonomy.
- Trained PyTorch detection and classification models for scratches, burrs, coating flaws and dimensional anomalies, tuned to the client's real accept/reject boundary — including acceptable cosmetic variation that had tripped up the off-the-shelf system.
- Deployed edge inference on NVIDIA Jetson devices at each station, classifying parts in milliseconds at line speed with no cloud dependency on the shop floor.
- Built a QC dashboard in Grafana showing per-line defect rates, defect-type Pareto charts and model-confidence trends, so quality engineers can spot upstream process drift early.
- Established a golden set and retraining loop: borderline images are routed to inspectors, their verdicts feed periodic retraining, and every model release is validated against the golden set before deployment.
Inspectors were not removed from the loop — they moved from inspecting every part to reviewing flagged parts and auditing samples, which is where their expertise actually pays off.
The Results
- 99.2% defect-detection accuracy against the client's golden set, with false-reject rates low enough that the line no longer stops for cosmetic non-issues.
- 70% fewer manual inspection hours, redeployed to flagged-part review, sample audits and upstream process work.
- 3x throughput at the inspection stage — the former bottleneck now keeps pace with the fastest lines.
- The project reached payback in about nine months, measured against the avoided third-shift staffing plan and reduced scrap and rework.
- The QC dashboard has caught two upstream process drifts before they produced customer-visible defects.
Figures reflect outcomes reported for this engagement; they are project results, not audited benchmarks.
Tech Stack
Services Used
Software Outsourcing — an AI/computer-vision delivery team owning the system from data collection to factory-floor deployment, on transparent pricing.