I will train a custom object detector for your classes with onnx and an API

Generic detectors miss the things your business cares about. I train an object detector on your own classes (products, defects, equipment, safety gear, anything you can label) and prove how well it works before you rely on it.

By package, you get a fine-tuned detector built on a model with a permissive license, exported to ONNX so it runs on CPU or GPU, a small inference API, and a clear report: per-class AP with 95% confidence intervals on a held-out test set, a comparison against an off-the-shelf zero-shot detector, and real examples of its mistakes.

In my sample project (hard hats, reflective vests and heads without a helmet), mAP@50 went from 22.3% with the generic detector to 57.3% after fine-tuning, measured on 1,584 held-out test images. The ONNX export matched the original model on the whole test set. It is a concept project, so your results will depend on your images and labels.

I check data quality first: duplicates, leaky train/test splits and unclear labels are the usual reasons a detector looks great in a demo and then fails in real use.

Please send labeled images in COCO, YOLO or Pascal VOC format, or tell me what you have and we will plan the labeling.