About

I built a system that predicts whether a Zerve user will become a Power User based on their early behavior, and tells the product team what to do about it.

The pipeline processes 409,000 platform events, builds behavioral features per user, and trains a classifier that achieves 96% accuracy. SHAP analysis explains every prediction — not just what tier a user will reach, but why.

The real problem this solves: platforms find out users churned after it already happened. This system catches them in week one. The biggest finding was simple — users who start an AI chat in their first week are 6.8 times more likely to become Power Users. So the API doesn't just classify users, it fires a specific recommendation: start that first chat, explore more features, or introduce the SDK.