Captured from the live deployment, not a mockup
01 · The problem
A detector that outputs only a label is unusable as evidence, so built a three-backbone ensemble of ResNet50, VGG16 and EfficientNetV2-B0 with calibrated confidence and an explicit borderline warning band.
02 · How it works
- 01
Trained on a two-stage schedule and served per-backbone Grad-CAM, so a reviewer sees which region drove each model's vote instead of trusting the aggregate score.
03 · What it cost, and what it returned
Published the training, evaluation and robustness evidence behind every number in an eight-page dashboard, and stated the scope limits plainly: face crops only, swaps rather than generated images, and a single dataset.