CASE STUDY DETAILS

Churn Guard

Project Type

Full-Stack ML Platform

Domain

FinTech / Banking

Challenge

The retention intelligence gap

Impact

operational Integrations in one workflow.

OVERVIEW

ChurnGuard is a full-stack machine learning platform designed to help banks and FinTech organizations identify customers who are at risk of leaving, understand the reasons behind their risk, and take proactive retention actions.

The platform combines a complete ML pipeline with an enterprise web application. Customer data is processed through data validation and feature engineering, after which an ensemble of machine learning models generates a churn-risk probability for each customer. The system uses SHAP explainability to identify the key factors contributing to each prediction, allowing analysts and retention teams to understand why a customer has been classified as high risk.

Beyond prediction, ChurnGuard converts risk insights into actionable retention strategies using Gemini AI. Analysts can review customer profiles, risk scores, revenue-at-risk, geographic and demographic trends, model performance, and SHAP-based risk drivers through an interactive Next.js dashboard. AI-generated retention recommendations can then be connected to operational workflows through integrations such as Jira and HubSpot.

PROBLEM STATEMENT

The source case
study identifies the
following operational problems.

The retention intelligence gap FinTech organisations need earlier, explainable signals for customers who may disengage. The source case study identifies the following operational problems

  • Recurring revenue is lost to silent customer churn
  • Retention teams discover disengagement only after customers leave
  • Generic outreach campaigns have low conversion rates
  • Risk flags lack explainable drivers that teams can act on
  • Customer data is scattered across multiple systems
  • Manual identification of at-risk customers is time-intensive and inaccurate

SOLUTIONS

1.  Proactive Churn Intelligence

  • Calibrated probability outputs for individual customer risk.
  • F1-optimised thresholding for practical classification
  • Batch reanalysis and 100K+ record CSV ingestion

3. AI Retention Strategies

  • Gemini 2.5 Flash generates personalised retention plans
  • Strategies are grounded in customer profile data and SHAP risk factors
  • Jira and HubSpot workflows connect insights to follow-up actions

5. Unified Analyst Experience

  • Next.js dashboard with risk distribution, geography, salary bands, revenue-at-risk, SHAP drivers, and cohort views
  • Customer explorer with drill-down analysis and per-customer strategy generation

2. Explainable Risk Decisions

  • Native TreeExplainer on XGBoost, LightGBM, RandomForest, and CatBoost
  • Exact ensemble SHAP values computed by linear averaging
  • Customer-level factors surfaced alongside each risk score

4. Enterprise Platform Controls

  • Multi-tenant organisation-scoped data isolation
  • RBAC for Super Admin, Org Admin, Data Analyst, and Viewer
  • JWT authentication, audit logging, API rate limiting, and real-time WebSockets

APPROACH

Our Approach

Data Engineering & Validation

  • Schema validation checks column presence, types, and range constraints
  • Chunked SQLAlchemy Core inserts support 100K+ record uploads

Feature Engineering & Store

  • Centralised ChurnFeatureRegistry maps API columns to the model schema
  • Domain features: Balance-to-Salary Ratio, Tenure-to-Age Ratio, Credit Utilization Proxy, and Engagement Score

Model Training & Selection

  • Optuna hyperparameter optimisation across 50 trials with 5-Fold Stratified CV
  • Calibrated Soft Voting ensemble: XGBoost + LightGBM + RandomForest + CatBoost
  • Decision threshold tuned for maximum F1

Explain ability Layer

  • TreeExplainer is applied to all four base estimators
  • SHAP matrices are averaged to obtain ensemble explanations without KernelExplainer overhead
  • Global feature importance is persisted to JSON

API & Backend

  • FastAPI service with 28 versioned endpoints under /api/v1/
  • WT authentication, RBAC, multi-tenancy, audit logs, WebSocket event bus, and SlowAPI rate limiting

    SUCCESS AND IMPACT

    What Was Achieved

    • Machine Learning Pipeline Development
    • Predictive Analytics & SHAP Explainability
    • AI-Powered Retention Strategy Generation
    • Multi-Tenant Dashboard Engineering
    • CRM Workflow Integration (Jira / HubSpot)

    Expected Business Impact

    • FinTech Companies
    • Digital Banking Platforms
    • Customer Success & Retention Teams
    • Enterprise Revenue Operations
    • Risk & Analytics Functions
    Scroll to Top