How it works
Roast My Plate
This image understanding demo is powered by Tante Fawzia, an Egyptian grandmother with fifty years of cooking experience and zero tolerance for mediocre plating. Select a dish, submit it for judgment, and receive a personalized, culturally-rooted critique. Nothing is ever good enough.
Business applications
1
Image Categorization and Tagging
Vision models can automatically ingest, analyze, and organize massive visual datasets with structured tags. For example, a food delivery platform could automatically scan restaurant menus or user-uploaded reviews to tag dish types, visible ingredients, and dietary categories (e.g., "gluten-free," "dessert," "grilled"), instantly organizing large catalogs with minimal manual data entry.
2
Automated Quality Control
Businesses can deploy visual analysis to assess physical products. In food service or manufacturing, the AI can evaluate presentation, portion sizes, or packaging against brand standards, instantly flagging items that fall short of quality expectations.
High-level technical workflow
1
Image understanding & visual analysis
The selected plate photo is sent directly to the vision model, which analyzes the actual visual content like dish type, presentation quality, colors, plating, portions, and more to ground the roast in what it genuinely observes rather than generic food commentary.
Tante Fawzia's entire persona, her speech patterns, Egyptian Arabic interjections, emotional arc from shock to grudging acceptance, and habit of comparing everything to her own cooking, is defined purely through the system prompt, demonstrating prompt engineering used as character design.
3
Structured output enforcement
The response is constrained to a strict JSON schema, forcing the model to return a validated list of short dialogue lines within defined length and count limits, ensuring consistently clean, frontend-ready data on every call.
4
Backend validation & response shaping
The AI response is validated by Pydantic on the backend before being returned to the client, guaranteeing that the frontend always receives well-formed, safe output.
