This project focused on authenticity checks for Game Boy Advance cartridges. The goal was to reduce guesswork by turning common collector checks into a repeatable machine-learning workflow.
The pipeline started with dataset gathering from listing photos, followed by cleaning and relabeling passes to remove noisy examples and standardize classes. I then trained and iterated on a YOLOv8 model for label recognition and authenticity classification.
One technical challenge was inconsistent source imagery. Cropping differences, compression artifacts, and poor lighting made class boundaries fuzzy, so data cleaning and class balancing were critical before training delivered stable results.
A concrete result was a faster first-pass review process: the model could flag likely fake listings earlier, reducing manual back-and-forth before deciding whether a listing needed deeper inspection.