Manufacturing
Vision-based QC — fewer false rejects, faster lines.
We implemented a computer-vision quality control system across two production lines. Models run at the edge, pipeline data fuels continuous re-training, and supervisors get near-real-time OEE visibility.
Overview
Operators were over-rejecting good units and re-inspecting borderline cases. We deployed a vision system that flags defects with explainable overlays and collects image samples to improve future models. The result: fewer unnecessary stops and smoother flow.
- Edge-first inference with GPU acceleration
- Operator UI with overlays and quick accept/override
- Automated data curation → re-training → safe rollout
- OEE/ SPC dashboards for supervisors & quality
- Full audit trail and model version traceability
From camera to decision to monitoring & improvement loop.
Approach
Instrument line cameras, collect labeled sets, measure baseline false reject rate.
Augment, train, calibrate thresholds; human-in-the-loop QA.
Optimize to ONNX/TensorRT; health checks, fallbacks, offline buffers.
Head-up defect overlays, quick actions, feedback capture.
Drift alerts, shadow tests, canary rollouts on shifts.
Golden image, IaC and device fleet management across sites.
Tech stack
- Edge inference (ONNX/TensorRT), GPU/Jetson deployment
- Training: PyTorch, Albumentations, Weights & Biases
- Data pipeline: Kafka, S3/Lake, Delta/Parquet
- MLOps: MLflow, feature store, model registry
- Monitoring: Prometheus, Grafana, OpenTelemetry
- OEE dashboards, SPC alerts, traceable audit logs
Steady reduction in false rejects after model iterations; throughput lift with latency kept under 120 ms.
“Operators trust the overlays, supervisors trust the dashboards. Quality meetings are now about trends, not screenshots.”
Ready to reduce rejects?
We can pilot on one line within weeks, then scale across sites with confidence.