About

ZervePulse: User Retention Intelligence Platform
Built entirely within Zerve's AI-native environment

THE PROBLEM:
Which user behaviors predict long-term success on Zerve? I defined success as sustained retention — a user who keeps coming back found real value. My goal was to identify what successful users do differently, flag who is at risk, and build a system to act on it before it's too late.

WHAT I BUILT:
ZervePulse is a full-stack retention intelligence platform built exclusively through Zerve's AI agent using natural language prompts — no external tools, no manual coding. It analyzes 409,287 behavioral events across 5,410 users, predicts individual churn risk, and segments users into actionable personas in real time.


DATA & FEATURE ENGINEERING:
Zerve's Data Processing Pipeline reduced 107 raw columns to 74. I then engineered 23 custom behavioral features per user, including:

AI Adoption Index — Depth of AI agent engagement.
Churn Velocity Score — Rate of activity decline.
Session Depth Score — Quality and intensity of sessions.
Time-to-First-Agent — Speed of AI agent discovery.
Unique Canvases — Breadth of work produced (top SHAP predictor).


MODELING & EVALUATION:
Four models competed — HistGradientBoosting won, evaluated across 22 metrics:

ROC-AUC: 0.7814
PR-AUC: 0.8622
F1 Score: 0.7490
MCC: 0.4136
Brier Score: 0.1865
Decile Lift: 1.54x

Validation also covered: SHAP explainability, DAU/MAU stickiness, Day 1/7/30 retention curves, CLV by persona, and revenue recovery estimates.


TECHNICAL CHALLENGES:
When XGBoost, LightGBM, imblearn, and SHAP were unavailable, the AI agent autonomously pivoted to sklearn equivalents and implemented a manual SMOTE — resolving NaN issues, dimension mismatches, and empty onboarding buckets independently. Every decision is documented and fully reproducible.


KEY FINDINGS:
1. Canvas Creation — the #1 Retention Predictor:
Unique canvas creation topped SHAP analysis (importance: 0.1322) — users who build more, stay more.

2. AI Agent Adoption — a Churn Firewall:
Users who engage with the AI agent churn at 52.0% vs. 77.1% for non-adopters — a 25.1 percentage point gap. It's the single most predictive success behavior on the platform.

3. Onboarding — the Critical Bottleneck:
Only 4.25% of users complete onboarding — the most critical activation gap I identified.

4. Five User Personas validated via ANOVA (F=1,392, p<0.0001):

Champions: 4.5% churn Deep AI adopters, high canvas creation.
Explorers: 87.6% churn Active but never finding a sticky workflow.
At-Risk: 87.5% churn Sporadic, clearly disengaging.
Ghosts: 55.7% churn Minimal activity.
Casual: 33.8% churn Moderate, stable usage.

5. Business Impact:
1,625 high-risk users flagged — at a 20% intervention rate, approximately 432 users are recoverable from churn, with direct revenue impact.


DEPLOYMENT:
Live and production-ready, with real-time user lookup, full leaderboard, persona filtering, and CSV export.

Live Dashboard: https://zervepulse.hub.zerve.cloud
Canvas: https://app.zerve.ai/notebook/a15ae300-6a9d-4609-a160-72fb0da3f594
Gallery: https://www.zerve.ai/gallery/a15ae300-6a9d-4609-a160-72fb0da3f594
GitHub: https://github.com/niloydebbarmacpscr/zervepulse