Alt text drafts for uploaded images

Design a feature for a blogging platform that drafts alt text for every image an author uploads, using a hosted vision-language model, so posts are more accessible to screen reader users. Authors upload about 30,000 images a day and should see a suggested description in the image settings panel, which they can accept or edit before publishing. The platform also has 4 million existing images with no alt text that should be backfilled over about four weeks without slowing down the live feature. Describe the flow for new uploads, the backfill, and what happens when the model output is wrong or unavailable.