What I built
- Benchmarked 7 CNN and transformer architectures on 27,560 images across 18 e-waste classes and selected ConvNeXt-Tiny.
- Developed a decision-support dashboard with 6 workspaces and 3 operational modes for single-image classification, mixed-scene review, and sampled conveyor belt video analysis.
- Implemented a hazard-aware policy layer that maps predictions to hazard level, material profile, disposal pathway, SDG 12 target, and human-review requirements.
Results
The selected ConvNeXt-Tiny model achieved 98.11% test accuracy and a 0.9811 macro F1-score across 18 e-waste classes.