Problem
A luxury brand needed a recommendation system, but standard collaborative filtering breaks for luxury — purchase frequency is low, statistical co-occurrence between customers is sparse, and brand aesthetic matters far more than price clustering. A customer who buys one piece does not follow the same pattern as the next.
Approach
Multimodal recommendation combining product imagery embeddings, structured catalogue metadata (collection, season, aesthetic tags), and natural-language descriptions. LLM layer generates the recommendation rationale ("this complements your earlier piece by carrying the same colour family from the autumn collection"), so customers see why the recommendation was made — important for a category where trust and storytelling matter as much as the suggestion itself.
Stack
Multimodal embeddings · vector index · Claude / GPT-4 for rationale · structured catalogue API · Python service layer
Outcome
Recommendations that respect the brand's aesthetic DNA rather than chasing pure statistical co-occurrence. The natural-language rationale layer outperformed silent recommendations on engagement — luxury customers expect a story, not a black-box suggestion.