RotoAnnotate

The annotation workflow combines Qwen vision-language proposals with human review on a canvas. Segmentation assistance and correction history connect labelling work to supervised fine-tuning and preference-based training.

A model proposal is a starting point, not an accepted label. Human corrections form training material, and held-out precision, recall and overlap checks gate a new adapter before promotion.

FastAPI execution and n8n directives organise the workflow. The architecture describes an evaluation loop; it does not prove an improvement in model accuracy without measured experiments.