Give the Feed Eyes

The situation

Two requests landed together. Creators want alt text generated for the images they post, so their feed is accessible without them writing a description every time. And the partnerships team wants the key terms — sponsor, fee, deliverables, dates — pulled automatically from the sponsorship PDFs creators upload, instead of a person retyping them. Both are 'the model looks at media and produces structured text'. Both also have a way of inventing details that aren't there.

What you'll learn

• AI-901 vision and extraction implementation: image description and alt text, structured extraction from documents against a schema, handling a schema mismatch, grounding output in the source media, and recognising fabricated detail.
• Object detection versus image classification versus OCR, image generation, and Azure Content Understanding: analyzers, extracted, classified and generated fields, confidence and grounding, and JSON versus Markdown output across documents, images, audio and video.

Who you work with

• 

  Jax Rivera

  Creator Partnerships Lead

• 

  Maya Chen

  Senior AI Engineer · Your mentor

Scenes

1. 1.Alt text and contracts
2. 2.What the image actually shows
3. 3.The vision and extraction clients
4. 4.When the schema doesn't fit
5. 5.Which analyzer?
6. 6.Vision and extraction check

What you leave with

The vision client generates alt text grounded in the confident object labels — not the model's invented colours and objects. The extraction client pulls the sponsorship terms into the payments schema, and when a contract has only a term in months rather than an end date, it catches the schema-mismatch error and falls back rather than letting the model guess a date. Every record links back to its source document.