Capability 08 · Vision AI

The demo worked.
The line is harder.

A deployment flow for vision models that need to keep working on a real line — as lighting shifts, dust settles, and edge cases arrive daily.

Edge deploymentLocal processingMonitoringRetraining
02

The situation

Accuracy in a demo and accuracy in production are different problems.

The model worked on clean test images. On the floor, the lighting changes through the day, dust settles on the lens, and new edge cases appear.

The hardware is limited, the network is unreliable, and results that arrive too late can’t be acted on.

A model that was accurate last month can quietly drift without anyone noticing.

03

The challenge

Staying accurate is the hard part — and noticing when you're not.

Getting a model to production is one thing. Keeping it accurate as conditions change is another.

It has to run reliably on edge hardware, feed results back fast enough to matter, and be watched so drift is caught early.

Without monitoring, the first sign of trouble is a missed defect or a storm of false alarms.

04

The capability

We take the model to the floor — and keep it accurate.

The model runs on edge hardware, checks the visual input locally, and sends structured results back into the operational system in real time.

Performance is monitored over time, so drift is caught before it becomes a problem, and the model is retrained as conditions change.

Inspection becomes part of the loop, not a separate step that ages out.

05

How the flow works

01The model is prepared and optimised for the edge
02It runs locally on the device, near the line
03Visual input is checked in place, in real time
04Structured results return to the operational system
05Uncertain cases are routed for review
06Performance is monitored for drift
07Alerts fire when accuracy slips
08The model is retrained as conditions change
06

Before / after

Before

A model that worked in a demo struggles on the line. Results are delayed, and drift goes unnoticed until something is missed.

After

The model runs at the edge, reports back in real time, and is monitored so drift is caught early and corrected. It keeps working as conditions change, not just at launch.

07

Controls

  • Local processing on the edge
  • Results logged and returned to the system
  • A confidence threshold for review
  • Human review for uncertain cases
  • Performance monitored for drift
  • Alerts when accuracy slips
  • Retraining as conditions change
  • A clear owner for the deployed model
08

Where this applies

Production-line inspectionEdge AI deploymentCamera-based workflowsQuality assurance loopsMeasurement supportReal-time alerts
09

What changes

Accuracy that holds in production
Results in real time
Drift caught early
Inspection in the loop
Fewer false alarms
A model that lasts
10

Questions about vision in production

What happens when the system is not sure?

It asks. Uncertain cases go to a person to review and approve instead of being decided automatically. The line keeps moving and the judgement calls stay with your team.