Brokers pasted WhatsApp text and phone photos, and that went live

Agriva's brokers are good at land and bad at copywriting. A listing wizard that asked for twelve fields in English got abandoned or filled with pasted chat messages.

Earlier drafts generated copy per language with separate calls, which cost more and produced versions that disagreed with each other.

• ✦One GPT-4o Vision call across the top photos that returns a structured list of what is actually visible: water bodies, fencing, road access, crops, buildings. The prompt forbids speculation.
• ✦One structured call that produces the title, highlights, SEO meta and descriptions in every requested language at once, from the broker's facts plus the visible features.
• ✦An amenities extractor that only confirms amenities present in the data or the photos.
• ✦A broker bio generator for the public broker page.
• ✦The broker reviews and edits before publishing; nothing goes live unseen.

• ✦Photos are ranked, the top set goes to the vision model, and its output becomes an input fact sheet.
• ✦Fact sheet plus broker text go into a single JSON-returning call keyed by language code, so every language is generated from the same facts.
• ✦Output is validated against the schema before it reaches the form.