Manufacturing QC · Human-in-the-loop
Inspect a case
Drop a ZIP of 65 sequential BMP frames. The model returns OK / Defective with confidence and Grad-CAM heatmaps on the most influential frames. You confirm or override; the decision is logged.
Input
No case selected
Reviewer stats
Reviews
—
Agreement
—
Overrides
—
Model
ArchitectureEfficientNet-B0 + Attention
Input16 × 224 × 224
Threshold0.48
Held-out accuracy84.6%
F1 / AUC-ROC0.857 · 0.839
Awaiting input
No case loaded
Upload a ZIP. The verdict, attention timeline across the 16 sampled frames, and the Grad-CAM overlays for the top-3 attention frames will appear here.
AI verdict
—
—
—
—
Attention across 16 sampled frames
Bar height shows the temporal attention weight assigned to each sampled frame. Higher bars = the model relied on that frame more.
Top attention frames · Grad-CAM
Human review
Confirm the AI's verdict or override it. The decision is logged for traceability.