A product photo contains useful clues about colour, category, and visible features. It does not contain everything a catalogue needs: warranty, dimensions, condition, included accessories, or a verified price.
SyntaxLab's product-intelligence demo analyses a photo, returns structured attributes, and drafts listing content in English and Arabic. When a seller supplies facts, those can be used to shape the generated listing. The workflow streams the analysis, applies image moderation, and stores demo records with a limited retention period.
Separate visible evidence from seller facts
The model may describe what it can see, but it should mark uncertainty and avoid asserting unseen details. A seller's structured data should remain the source for price, stock, warranty, and condition. If an attribute is inferred from an image, present it for review.
This makes the interface useful without silently turning a visual guess into a product specification.
Build a reviewable catalogue record
Return fields that the commerce system can validate: title, category, attributes, description, language, and image reference. Let the seller edit the output before export or publication. Keep the original image and the edited facts connected so later updates can be audited.
Image models can analyze objects and visual attributes, but image quality, angle, lighting, occlusion, and model uncertainty still affect the result. OpenAI Images and Vision
Evaluate by product category and language
Measure attribute accuracy separately across categories, photo styles, and languages. Arabic copy needs a fluent review for terminology, dialect, and right-to-left display. A fluent translation can still be wrong about the item.
Project evidence
The demo is implemented in syntaxlab-backend/src/products.ts and surfaced in syntaxlab/src/routes/demos.product.tsx. It analyses uploaded product images and can generate bilingual listing content. The repository does not show a live commerce catalogue, publication integration, or measured listing quality.