The challenge
The client is a general merchandise retailer with both physical stores and a substantial e-commerce presence, sitting on years of point-of-sale, web, and app data that had never been unified into a single customer view. However, they faced significant hurdles:
- Data Silos: Online and offline purchase data were never merged, so the same customer looked like two different people to the business.
- Generic UX: Every visitor saw the same homepage products regardless of browsing or purchase history.
- Slow API: The existing search stack was too slow for real-time suggestions at storefront-scale traffic.
- Cold-Start Problem: New shoppers with no purchase history got no useful recommendations at all under the old rules-based system.
Our solution
We built a custom recommendation engine on top of a newly unified customer data platform:
- Data Lakehouse: Consolidated data from point-of-sale, website, and app into Databricks, creating a single customer view for the first time.
- Collaborative Filtering Models: To suggest "customers who bought X also bought Y" using purchase-pattern similarity.
- Real-Time API: Served recommendations in under 50ms during page loads, fast enough to appear above the fold without a layout shift.
- Cold-Start Fallback: A content-based model kicked in automatically for new visitors until enough behavioral signal was collected.
Conversion Rate by Segment
Implementation roadmap
Building pipelines from 6 core data sources into the unified Lakehouse.
Training and back-testing collaborative filtering and cold-start models against historical purchase data.
Testing personalized recommendations against the generic homepage across a meaningful traffic split.
Retraining models weekly as new purchase and browsing data accumulates.
Key results
Personalized product suggestions drove an 8% increase in conversion rate. The "Recommended for You" section became one of the highest-revenue placements on the site, outperforming several long-standing merchandised banners.
The unified customer data platform also became a foundation the marketing team now uses for email segmentation and loyalty program targeting, well beyond the original recommendation-engine scope.
"We finally feel like we know our customers. The AI suggests products they want before they even know they want them."