Real demo recording · English narration and captions · Fictional data
THE PROBLEM
Visual inspection needs more than an alert: a normal reference, evidence of the change and a traceable decision.
THE WORKFLOW
Compare synthetic panels against a normal-surface memory bank. A local ResNet18 produces anomaly distances; the reviewer inspects the heat map, records a decision and exports the evidence.
01Normal-surface comparison
02Anomaly distance and heat map
03Human decision with revision
04Report and export
RECORDED EVIDENCE
Results, with context.
Results on synthetic data
Check
Cases
Test panels: correct classification
16/16
Interface examples: correct classification
4/4
Dataset
original-controlled-fixtures-v1
Completed cases
20/20
Median / p95 per case
2.81 s / 65.306 s
Evaluation record (UTC)
2026-10-08
Source: the project’s evaluation record, using local inference. Recorded hardware: NVIDIA GeForce RTX 4070 Laptop GPU / 8 GB. Latency includes the full HTTP case and may include initial model loading. These measurements have not been rerun on the machine serving this website.
An inspection pilot with normal references, human review and a separate evaluation set.
What we would define first
Part types, tolerances, normal references, capture conditions and the cost of false positives and negatives.
Demonstration limits
Sixteen test panels use seeds separate from the references, but the same synthetic generator. This does not measure industrial accuracy. The heat map is not certified defect segmentation.
Evaluated demo
Evaluated demo