Overview
Designed and deployed a machine learning-powered recommendation engine integrated with a scalable microservices e-commerce platform on GCP. The solution delivered personalized product suggestions using real-time behavior data, increasing conversion rates while reducing bounce rates.
17%
Increase in conversion rate
30%
Improvement in average basket size
Key Components
- User interaction and order data streamed via Pub/Sub
- Data transformation and enrichment via Cloud Dataflow
- Feature extraction and scoring done in Vertex AI Pipelines
- Redis used for real-time inference and caching
- Model training and deployment automated with Vertex AI Training/Prediction
- Monitoring with custom dashboards to track performance and drift
Architecture Strategy
- Seamlessly embedded ML into API responses using service adapters
- Delivered predictions under 200ms via Redis lookup and fallback scoring
- Batch and streaming pipelines co-existed for cold-start and long-tail coverage
- Continuous evaluation loops with A/B testing setup
Architecture Flow: User events → Pub/Sub → Dataflow → Vertex AI → Redis → Recommendations API
Engineering Considerations
- Built CI/CD pipelines with Cloud Build and Artifact Registry
- Feature engineering pipelines modularized for model reuse
- Used BigQuery for offline evaluation and feature logging
Business Outcomes
- 17% lift in click-through and checkout conversion
- 30% improvement in average basket size for users with personalized flows
- Enabled business teams to tune recommendation weights using config-as-code
Why It Worked
- Event-driven, decoupled AI pipelines
- Real-time inference embedded without blocking API latency
- Integrated model governance and transparency