Kestrel Industrial

Mobile application

The challenge

Kestrel inspected roughly 4% of output through statistical sampling — the industry norm, and adequate until an OEM customer found a defect pattern that sampling had structurally missed for eleven weeks. The resulting recall cost $4.7M and put a major program at risk. Scaling inspection meant hiring inspectors at a rate the labor market couldn't supply, and the machine data that might have predicted the tooling wear behind the defect sat locked inside proprietary controllers.

Research

• Protocol and connectivity survey across all nine plants and 340 machine tools
• Root-cause analysis of the recall event traced to progressive tool wear
• Time-and-motion study of manual inspection to establish the accuracy baseline
• Historical failure analysis across four years of maintenance records

The solution

Edge gateways read from controllers over OPC-UA and Modbus in a strictly read-only posture on a segmented network — production control was never placed in the data path, a non-negotiable for the plant safety review. Vision inspection runs at the line with models trained per part family; ambiguous classifications route to a human review queue rather than being forced to a decision, which both preserved quality and generated continuous labeled training data. Condition-based maintenance models consume vibration and thermal telemetry to flag tool wear before it produces defects.

UX decisions

Features

• Read-only OPC-UA and Modbus gateways on segmented OT networks
• Per-part-family vision inspection at 100% coverage
• Human-in-the-loop review queue feeding continuous retraining
• Condition-based maintenance from vibration and thermal signals
• Line-level OEE tracking with downtime attribution
• ISO 9001 and PPAP documentation generation
• Defect Pareto analysis by line, shift, tool, and program

Technology

Architecture

Edge gateways run containerized protocol adapters, publishing normalized telemetry to Kafka over a one-way network path from OT to IT — there is no route back into the control network by design. Vision inference runs on edge GPUs at each line, with only classifications and flagged images crossing to the cloud, keeping bandwidth manageable across nine plants. PyTorch models are versioned and deployed through an MLOps pipeline requiring a validation gate against a per-part-family holdout set. Go services aggregate plant state; the whole cloud estate runs on AKS.

Results

Inspection coverage went from 4% to 100%, and detection caught 3.2× more defects than the manual baseline on parallel-run comparison. Escape rate — defects reaching the customer — fell 87%. Condition-based maintenance cut unplanned downtime 34%. The tool-wear pattern behind the original recall is now caught, on average, 9 days before it produces an out-of-tolerance part.

Lessons learned

1. 01The read-only OT posture cost us three months of additional engineering and was the reason the plant safety review passed. There was no version of this project that survived putting inference in the control path.
2. 02The human review queue was proposed as a temporary bootstrap and became permanent infrastructure. Uncertainty routing is a feature, not a training-phase crutch.
3. 03Per-part-family models substantially outperformed a single general model, at the cost of a much heavier MLOps burden. The pipeline investment was what made that tractable.