Back to Case Studies
AI/ML

AI-Powered Recommendation Engine

Increasing user engagement by 45% through personalized recommendations

Recommendation SystemsPersonalizationB2CML
6 months (MVP to production)
Team of 12
Lead Product Manager

🎯The Problem

Users were struggling to discover relevant content in a catalog of 100K+ items. Manual curation wasn't scalable, and generic trending lists had low click-through rates (2.3%). Customer surveys showed 68% of users felt overwhelmed by choices, leading to decision paralysis and churn.

🔍My Approach

I led a cross-functional initiative to build a personalized recommendation system: 1. **Data Strategy**: Collaborated with data science to identify key signals (views, clicks, purchases, ratings, session duration) 2. **ML Approach**: Evaluated collaborative filtering vs content-based vs hybrid approaches 3. **User Research**: Conducted 25 interviews to understand discovery patterns and preferences 4. **Experimentation Framework**: Designed A/B tests with clear success metrics (CTR, engagement time, conversion) 5. **Phased Rollout**: Started with 5% traffic, scaled to 100% after validation

💡The Solution

Built a hybrid recommendation system combining collaborative filtering and content-based approaches: **Phase 1 - MVP (3 months)** - Implemented collaborative filtering using user-item interactions - Created "You Might Like" section on homepage - Built real-time inference API with <100ms latency **Phase 2 - Enhancement (2 months)** - Added content-based features (categories, attributes, descriptions) - Implemented contextual recommendations (time of day, device, location) - Created personalized email campaigns **Phase 3 - Optimization (ongoing)** - A/B tested different model architectures - Added explanations ("Because you liked X") - Implemented diversity constraints to avoid filter bubbles

Technologies Used

PythonTensorFlowAWS SagemakerRedisSparkMLflowLooker

📈Impact & Results

+187%
Click-Through Rate
From 2.3% to 6.6% on recommended items
+45%
User Engagement
Average session duration increased from 8 to 11.6 minutes
+32%
Conversion Rate
Users who engaged with recommendations converted at 32% higher rate
$12M
Revenue Impact
Incremental annual revenue from personalized recommendations
+28%
User Satisfaction
NPS score improved from 42 to 54

💭Key Learnings

  • •Cold start problem was harder than expected - invested in hybrid approach combining popularity with content similarity
  • •Explainability matters - adding "Why this?" increased trust and CTR by 15%
  • •Diversity vs relevance tradeoff required careful tuning - users wanted some serendipity
  • •Real-time vs batch predictions - chose batch with 15-min refresh for cost optimization
  • •Model monitoring essential - set up alerts for CTR drops, coverage issues, and latency spikes

Want to Learn More?

I'd be happy to discuss this project in more detail, share additional insights, or answer any questions.

Let's Connect