Client churn model at Wakilni

The problem

Wakilni earns per delivery, so no client ever cancels; they stop sending orders, and by the time someone notices the revenue is gone. Churn wasn't a column in the data; defining it from each client's own ordering rhythm was most of the work.

What I built

• XGBoost with calibrated probabilities and SHAP explanations; Kaplan-Meier and Cox survival analysis.
• FastAPI service, scheduled scoring pipeline, dashboard sorted by revenue at risk; deployed on a hosted server.
• Designed the company's shared AI-platform repository (gateway, scheduler, one container per model); one of two contributors chosen.