AltTextLab

A retailer with tens of thousands of listings has alt text on almost none of them. A scheduled agent queries Shopify for products whose images have an empty alt attribute, and for each one calls Generate Alt Text with the product title, brand, color and material attached so the model runs in e-commerce mode, plus the collection name as page context. Custom fields carry the product ID and image ID through the call and come back unchanged, so the agent writes each result to the right image without a second lookup. Descriptions the model produced from the image alone are written directly. Descriptions that depended on the supplied context are queued for review.

An editorial agent watches for posts moving to Ready to Publish in WordPress. For each embedded image it assembles the page title, the nearest section heading and the surrounding paragraph, then calls Generate Alt Text in the Neutral style with the post's target keywords supplied. The descriptions are written into the post before it goes live, so the accessibility work happens at publication rather than in a remediation project a year later. If any image URL fails to resolve without a redirect, the post is held and the reason is reported, because a silently skipped image is exactly how a backlog rebuilds itself.

A site published in five languages has alt text in one. Rather than translating existing strings, which propagates whatever was wrong in the original, the agent calls Generate Alt Text once per language directly against the image, selecting the language by name from the supported list and passing the localized page title and section heading as context. Each locale gets a description written from the image itself, and the results are written into the matching entries in Contentful. Failed responses cost nothing, so a language that returns errors can be retried without spending credits on the attempt.