AI Fashion Discovery Platform
Multimodal Search & Personalized Ranking — A mobile-first fashion discovery platform where shoppers search a catalog of nearly a million products by natural language, reference photos, or both together, and get routed to the right item on the original brand's site.
What we were solving for
Shoppers describe what they want in very different ways — a mood, an occasion, a brand, or a reference photo. Keyword search can't read aesthetics or visual similarity, and vector search alone can't guarantee exact matches on brand, size, or price.
The catalog behind it was just as fragmented, spanning over a thousand brands with different schemas, currencies, and description styles for the same silhouettes.
How we built it
Normalized multi-brand catalog
Products from over a thousand active brands are continuously ingested and normalized into one shared schema.
LLM-based attribute enrichment
Structured attributes — silhouette, occasion, formality, style — are derived from product text and images, replacing inconsistent brand copy.
Query-to-product-description search
An LLM rewrites a shopper's query as a hypothetical product description before embedding it, combined with exact SQL filters on brand, size, price, and color.
Controlled filter relaxation
When strict filters return too few results, lower-priority constraints relax gradually, and the response reports exactly what was widened.
Combined image + text search
A photo with added text is classified as a substitute, refinement, occasion, or mood request, so each signal guides a different part of the result.
Personalized, explainable ranking
A style profile built from interaction history informs ranking, with results explained in plain language proportional to actual confidence.
Where it landed
Launched as a live, mobile-first discovery product searching across nearly a million garments from over a thousand active brands.
Combines natural-language, visual, and combined image-and-text search in one retrieval system.
Fixed a 245× vector-search latency cliff, taking warm searches from 10+ seconds down to 14–34ms as the index grew roughly 7.5×.
More Engineering Work
View All Case Studies →Have a similar system to build or rescue?
Discuss your technical requirements directly with our lead product engineers.
Schedule a Scoping Call