About

Which Zerve Users Stay and Why
A full behavioral analysis of 409,000 events across 6,158 users to answer one question: what separates the 2.6% of users who become long-term power users from everyone else?
This notebook walks through the complete analysis, from raw PostHog event data to machine learning models, including data cleaning, user-level aggregation, day-zero behavior analysis, onboarding funnel exploration, and a blind prediction test run on Zerve's own AI agent before the models were trained.
Key findings: retained users open a canvas and engage the agent within their first session, reach 20 meaningful actions faster, and combine agent usage with manual exploration at a 159x higher rate than churned users. Onboarding form completion is a false signal. The credit wall is not the primary churn driver. The most powerful feature on the platform is the one most users never meaningfully touch.
Built entirely on Zerve using Python, pandas, scikit-learn, and matplotlib. Models include Logistic Regression and Random Forest with balanced class weights.