AI Data Labeling
I led a bounding box annotation project for 50,000 retail shelf images, producing ~400,000 boxes across 1–15 objects per image. The team of 20 labelers worked over 6 weeks. Quality was enforced via double-labeling 10% of images (requiring IoU ≥ 0.7), weekly audits of 5% of batches (precision/recall > 95%), and strict edge-case rules for occluded or blurry objects. A 2% holdout set was verified against human ground truth.