CMAB Recommendation Engine

Contextual multi-armed bandit recommendation engine with XGBoost, handling 5K requests/second for real-time ML predictions.

Overview

A contextual multi-armed bandit recommendation engine powering Optimizely’s experimentation platform with XGBoost, enabling personalized A/B test variations based on individual user context.

How It Works

  • Contextual Bandits personalize user experience by considering context-specific factors
  • XGBoost powers the prediction layer for real-time decision making
  • Large-scale data feed with regular interval training
  • Context-aware bandit system that reduces regret and improves experiment outcomes

Tech Stack

Python XGBoost FastAPI GKE BigQuery Datadog

Impact

  • Handling 5,000 requests/second for real-time ML predictions
  • Reduced regret by 15% and boosted conversions by 8%
  • Secured a major US financial services client with tens of millions of customers
  • Notable clients: Discovery, Atlassian, IBM, Salesforce, Capital One