Mehdi Sattari

I developed a comprehensive churn-prediction framework for iKhokha using a dual-strategy approach. The first component is an XGBoost-based ML model that predicts churn probability over a rolling 30-day window. The second is a process-control system known as "Triggers", designed to detect unusual shifts in merchant activity patterns that could be due to business changes, hardware failures, or unexpected shutdowns, allowing for quicker intervention in the event of unexpected churn.

The XGBoost model generates monthly scores for the entire active merchant base, while the "Triggers" system provides daily risk updates. I deployed the entire pipeline as Snowpark UDFs and Stored Procedures within Snowflake, ensuring high-performance, in-database execution and automated scheduling.

To drive business impact, I built interactive Power BI and Tableau dashboards that track predicted vs. actual churn, recovered revenue, and cohort-level trends. This allows stakeholders to monitor intervention effectiveness in real-time and make data-driven decisions.

I designed a retention-led A/B test validating a 31% uplift in merchant retention. This experiment proved that my dual-strategy framework significantly outperforms the traditional baseline.

At Elevate Credit, I delivered the fourth-generation Rise and Elastic subprime risk models, focusing on 6-month charge-off prediction. I used Boruta and Bayesian optimization (Optuna) to distill over 4,000 attributes from Clarity and Experian into fewer than 70 stable features, thereby enhancing both model efficiency and regulatory interpretability. To ensure these models were production-ready and compliant, I integrated monotonic constraints, SHAP explainability, and rigorous shock analyses. The outcome was restored portfolio performance and an AUC that rose from 0.58 to a robust 0.66. To manage the complexity of high-dimensional credit data, I overhauled the ML pipeline with multiprocessing and nested CV, accelerating feature selection by 80% and training by 40%.

The models were ultimately validated rigorously via smoke tests, then containerized and deployed via Azure Docker containers. To ensure long-term stability, I implemented automated drift detection and Microsoft Power BI dashboards for real-time PSI and AUC monitoring. This provided a stable, scalable foundation for the organization's strategy, ensuring the models remained robust against economic shifts while meeting strict regulatory standards.