StyleVault

StyleVault's text-based product search was failing customers who wanted to find items by visual similarity. Shoppers would screenshot outfits from social media but had no way to translate those images into product matches. This led to a 68% search abandonment rate and an estimated $4.2M in lost annual revenue from unfulfilled purchase intent.

Mixpeek's multimodal pipeline ingests every product image and extracts visual embeddings using CLIP-based feature extractors. A custom retriever enables shoppers to upload a photo or paste a URL and instantly receive visually similar products ranked by style, color, and silhouette. The system also cross-references text metadata to boost relevance when shoppers add optional keywords.

The integration began with a two-week proof of concept covering 200,000 SKUs. After validating retrieval precision above 92%, the team scaled to the full 2M-item catalog over the following month. A Celery-based batch pipeline re-indexes new products nightly, while Ray Serve handles real-time inference at sub-200ms p95 latency.