Local AI can put the inspection computation close to the production event. The engineering question is how that computation fits the station cycle, interfaces and operating requirements.

From a model to a working station.

EDGE AI FOUNDATION’s public “2026 and Beyond” overview identifies real-time quality control as a manufacturing application. Edge Impulse’s current visual inspection offering combines data capture, model development and management with defect detection and assembly verification. These are useful signals of a broader move toward deployable inspection workflows.

Choose compute for the workload.

Image resolution, camera count, inspection methods and cycle time determine the resources required. A smart camera, industrial computer or local accelerator system may each be appropriate. Include transfer, handling and controller communication when evaluating timing.

Make failures explicit.

The workflow should define what happens when an image is missing, a check is incomplete, product identity is lost or an interface is unavailable. Local inference belongs within that complete station behavior.

Use larger models for a defined task.

Edge Impulse hosts an expert demonstration combining visual anomaly detection with a vision-language model for contextual description. The project also reports incorrect descriptions and identification problems. Such demonstrations suggest directions for assisted review while reinforcing the need for bounded tasks and independent evaluation.

Manage the inspection over time.

The March 2026 VDA yellow-volume publication on AI in quality management discusses data, model risk and changing operating conditions. It is a non-binding draft recommendation, used here as an industry direction signal. Project obligations come from the applicable customer requirements and agreements.

For Reliova, edge AI is part of system engineering: a suitable local architecture, controlled deployment and results connected to the quality workflow.

Further reading

All engineering insights