The Challenge
A major North American e-commerce retailer with over 50 million active customers and a catalog of 2 million products was struggling with their personalization strategy. Despite collecting vast amounts of customer behavioral data, their recommendation system was producing generic, poorly-targeted suggestions that customers largely ignored.
The existing system relied on simple collaborative filtering that couldn't keep pace with their rapidly changing inventory and seasonal trends. Recommendations often featured out-of-stock items, ignored customer context (device, time, location), and failed to account for the complex interplay between online browsing and in-store purchases.
The business impact was significant: homepage personalization drove only 3% of revenue, and recommendation widgets had click-through rates below industry averages. The company estimated they were leaving $200M+ in annual revenue on the table due to poor personalization.
The Solution
Jasnova built a comprehensive personalization platform that unified customer data across all touchpoints and delivered real-time, context-aware recommendations at massive scale.
Technical Architecture
- Real-Time Customer 360: Unified streaming pipeline that combines online behavior, transaction history, loyalty data, and store visits into a real-time customer profile
- Multi-Stage Candidate Generation: Combined multiple retrieval strategies including collaborative filtering, content-based similarity, trending items, and business rules to generate diverse candidate sets
- Deep Learning Ranking: Transformer-based ranking model that scores candidates based on customer context, session intent, and business objectives
- Multi-Armed Bandit Optimization: Implemented Thompson Sampling for continuous optimization of recommendation strategies, balancing exploration and exploitation
Personalization Contexts
The system delivers optimized recommendations across multiple touchpoints:
- Homepage: Personalized hero content, category suggestions, and product carousels based on customer lifecycle stage and recent behavior
- Product Pages: "Frequently bought together," "Customers also viewed," and "Complete the look" recommendations
- Cart & Checkout: Last-minute upsells and cross-sells optimized for incremental revenue without cart abandonment
- Email: Personalized email content including abandoned cart recovery, replenishment reminders, and new arrival alerts
- Search: Re-ranked search results based on customer preferences and predicted purchase intent
Implementation
Phase 1: Data Foundation (Weeks 1-8)
Built the real-time customer data platform using Apache Kafka and Apache Spark. Integrated data from 15+ sources including web analytics, mobile app events, POS transactions, email engagement, and customer service interactions. Established customer identity resolution to unify profiles across devices and channels.
Phase 2: Model Development (Weeks 6-14)
Developed and trained the recommendation models using 18 months of historical data. Created a comprehensive feature store with over 1,000 features spanning customer demographics, behavioral patterns, product attributes, and contextual signals. Validated model performance through extensive offline evaluation and A/B testing frameworks.
Phase 3: Platform Build (Weeks 10-18)
Built the serving infrastructure to handle 10,000+ recommendations per second at sub-100ms latency. Implemented multi-level caching, model sharding, and graceful degradation strategies. Created APIs for each recommendation context with fallback strategies for cold-start users and new products.
Phase 4: Rollout & Optimization (Weeks 16-24)
Gradual rollout starting with homepage recommendations, expanding to other touchpoints based on measured lift. Continuous A/B testing of model variants and recommendation strategies. Integration with merchandising tools for manual boost/bury controls.
Results
After six months of production operation:
- 25% Increase in Conversion Rate: Customers who engaged with recommendations converted at 25% higher rates than those who didn't
- 15% Higher Average Order Value: Effective cross-sell and upsell recommendations drove larger basket sizes
- 40% of Revenue from Personalization: Revenue attributed to personalized recommendations grew from 3% to 40% of total online revenue
- 3x Email Click-Through Rates: Personalized email campaigns saw triple the engagement of generic campaigns
- 22% Reduction in Bounce Rate: Relevant homepage personalization kept customers engaged longer
Key Innovations
Session Intent Detection
The system detects real-time shopping intent from browsing patterns. A customer browsing multiple items in a category signals exploration mode (show variety), while repeated views of a specific item signal purchase intent (show alternatives and complementary items). This dynamic intent detection improved click-through rates by 35%.
Inventory-Aware Recommendations
Recommendations are continuously adjusted based on real-time inventory levels. Items with low stock are down-ranked to prevent customer disappointment, while overstocked items receive a boost to help clear inventory. This optimization generated an additional $15M in margin improvement by reducing markdowns.
"Jasnova's personalization engine has become the backbone of our customer experience. The system understands our customers better than we ever could with manual merchandising, and the results speak for themselves."
- Chief Digital Officer, E-commerce Retailer
