I train a custom object detection model on your labelled images with two Apache-2.0 detectors, RF-DETR Nano and YOLOX-Nano, instead of Ultralytics YOLO, which is AGPL-3.0 (its GitHub repo). You get both models, tested, with a report comparing them.

Both models as ONNX, tested in ONNX Runtime, plus the PyTorch weights (.pth)

A short Python script that runs either one on CPU

A report (HTML + PDF): per-class precision, recall, mAP50 and mAP50-95 on held-out images, failure galleries, CPU speed and a licence table with sources

A check of your labels first: broken boxes, duplicates, images copied across splits

One retrain after you fix labels the report flags, or a different confidence threshold

You send labelled images: COCO, YOLO txt, Pascal VOC or a Roboflow export, up to 2,000 images and 10 classes.

Not included: labelling, the training scripts themselves, installing on your device, face or person identification, licence plates, security cameras, and cloud GPU costs.

Your images stay on my computer. They never go to an AI service without your written yes.

The licence table states published licences; it is not legal advice. I report measured accuracy; I can’t promise a number before training.

AI use: AI tools run the scripts and draft the report from the numbers; they don’t see your images. I check every number myself.