← All works
Machine learning2026
[DRAFT] Churn Radar
Churn prediction pipeline for a subscription retailer — from feature store to live scoring and alerting.
Nordwind's retention team was flying blind: churn showed up in the quarterly numbers long after the customer had already gone quiet.
We built a feature store on their existing warehouse, trained a gradient-boosted churn model, and wired live scoring into their CRM so at-risk accounts surface the moment behaviour shifts.
The model retrains nightly and pushes a ranked at-risk list into the tools the retention team already lives in.
+34%
Churn caught early
<800ms
Scoring latency
Nightly
Retrain cadence
Stack
PythonPyTorchPostgreSQLAWS
Built for [DRAFT] Nordwind Retail