The Challenge
ShopTrend, a fast-fashion e-commerce retailer, had a massive catalog of over 50,000 SKUs. However, their users were overwhelmed. The basic “sort by popularity” and “sort by price” filters weren’t enough to help users find items they actually loved. The result was a high bounce rate and a low average order value.
They wanted to deliver a “personal stylist” experience at scale.
The Solution
We built RetailAI, a hyper-personalized recommendation engine that understands visual style, not just keywords.
Vector Search & Visual Similarity
We didn’t just match text tags. We used computer vision models to analyze product images, extracting feature vectors that represented “style,” “cut,” and “pattern.” Using a vector database, we could instantly show users “visually similar” items, even if they were tagged differently.
Real-Time User Intent Modeling
We tracked real-time user behavior—clicks, hovers, and dwell time. This session data was fed into a model that adjusted the homepage feed instantly. If a user lingered on red dresses, the feed would subtly shift to feature more red items and matching accessories.
Key Technologies
- Pinecone (Vector Database): For high-performance similarity search.
- OpenAI Embeddings: For semantic text search.
- Python (FastAPI): For the high-performance inference API.
- Next.js: For the server-side rendered storefront.
The Results
The “For You” feed quickly became the most visited section of the site.
- 45% Increase in Conversion Rate: Users found what they wanted faster.
- 20% Increase in Average Order Value: The “Complete the Look” recommendations drove cross-sells.
- 30% Reduction in Bounce Rate: The personalized homepage kept users exploring.
“It feels like magic. Our customers think we have a team of stylists curating the site for them, but it’s all the AI engine Sparkable built.” — Amanda Lee, Head of Product at ShopTrend