International Churn Model Deployment
2023 · AWS SageMaker, Snowflake, S3, XGBoost, GitLab CI/CD
Problem
A centralized churn model existed on the ML platform but couldn't be operationalized across markets. Each country had its own data warehouse schema and marketing automation system, so any deployment would require duplicating the model logic per country — high maintenance cost, high drift risk.
The challenge: deploy a single model artifact to 5 countries with heterogeneous tech stacks, enable A/B testing in each, and do it without the codebase fragmenting into country-specific forks.
Solution
Started with churn analysis to understand where the signal was strongest, then moved into feature engineering, model development, and sampling strategy for imbalanced datasets. Predictions are stored in the central ML platform and distributed to local entities via scheduled S3 jobs — each country reads from S3 into its own data warehouse and marketing automation system.
The deployment is model-agnostic: adding a country is a config change, not a code change.
Architecture
Impact & Scale
- First international ML model deployed across 5 countries
- Each country has run multiple A/B tests — some quarterly, some monthly, some weekly
- Setting up bi-weekly international check-in as a flywheel to systematically improve experiment cadence and results
Next Steps
- Increasing retention uplift modelling using causal inference (see uplift prediction project)
- Launching a self-service retention portal where customers can select and claim retention offers directly