Deriv Defensor - AI Partner Churn Prevention

THE PROBLEM Deriv's affiliate partner program faces a critical challenge: 20% annual churn rate costing $32 million in lost lifetime value. Partners leave silently—by the time relationship managers notice declining engagement, it's too late to intervene effectively. The business needed a system to predict churn early, understand WHY partners disengage, and trigger timely interventions. THE SOLUTION Deriv Defensor is an AI-powered churn prevention system that combines machine learning prediction, causal inference, and GenAI explanation into a unified early warning platform. Key Components: 1. PREDICTION ENGINE - LightGBM model trained on 10,000 partners across 18 behavioral features (login trends, revenue velocity, payment patterns, network health, support tickets). Achieves 99.4% AUC with robust cross-validation. 2. CAUSAL INTELLIGENCE - Uplift modeling identifies which interventions work for which partner types, discovering interaction patterns not explicitly programmed (e.g., declining revenue + declining commissions creates 67% synergy effect—4x baseline churn risk). 3. GenAI EXPLAINABILITY - Claude AI translates model outputs into human-readable explanations. Instead of "churn probability: 87%", managers see "This partner is 15 points from tier demotion. Performance anxiety detected. Recommend immediate call with tier protection offer." 4. AUTOMATED INTERVENTION - Real-time alerts trigger personalized outreach with drafted emails, intervention logging, and learning feedback loops. BUSINESS IMPACT - 5.1x better ROI vs random targeting - $25M incremental value protected - 172 partners saved (vs 127 random) - 26% lower cost per save ($116 vs $157) TECHNICAL ARCHITECTURE Production-ready stack: Python + LightGBM + Claude API + Streamlit + PostgreSQL. Features real-time monitoring, cross-validation (99.4% ± 0.2% AUC), noise robustness testing (89% expected production AUC), and continuous learning from intervention outcomes.