BankSight AI

Inspiration

Retail banking institutions across Europe lose billions of Euros annually to customer attrition. Traditional customer retention strategies suffer from two major flaws:

Opaque Black-Box Models: Legacy scoring mechanisms fail to explain why an account is departing, violating modern regulatory mandates under the EU AI Act, GDPR Article 22, and Basel III / European Central Bank (ECB) supervisory standards. Blind Resource Allocation: Traditional marketing departments target customers with high churn risk regardless of whether marketing intervention actually helps—often inadvertently provoking "Sleeping Dogs" (customers who churn specifically because they were contacted).

We engineered BankSight AI to transform retail banking retention from a reactive guessing game into an explainable, causal, and mathematically optimized enterprise intelligence system.

What it does

BankSight AI is a production-grade full-stack banking intelligence platform built on 10,004 audited European retail banking accounts across France, Spain, and Germany (€765.4M in total monitored balance exposure). The platform delivers 22 dedicated analytical modules:

• Interactive 3D Holographic Landing & Customer Galaxy: Features real-time 3D rotating holographic customer orbs and bank pediment emblems, coupled with an interactive Three.js WebGL 3D point-cloud projecting 10,004 customer embeddings across PCA dimensions.

• Multi-Model Benchmark Suite: Evaluates and calibrates 5 algorithms—Gradient Boosting (0.8612 ROC-AUC Champion), XGBoost, CatBoost, Random Forest, and Logistic Regression—with Brier calibration curves and PR-AUC metrics.

• Explainable AI via TreeSHAP: Provides global beeswarm rankings and customer-level waterfall attributions, revealing critical banking phenomena such as the Multi-Product Holding Cliff (where churn surges from 7.58% at 2 products to 82.71% at 3 products, p < 0.0001).

• Evidence-Aware Retention Allocation (EARA): A novel optimization framework that solves a constrained bounded knapsack problem under review budgets, balancing churn hazard, normalized capital balance, and cross-model uncertainty penalties.

• Two-Model Causal Uplift Engine: Estimates the Individual Treatment Effect (ITE), partitioning accounts into Persuadables, Sure Things, Lost Causes, and insulating Sleeping Dogs.

• Household Contagion Network: Force-directed relational graph mapping co-holder relationships and multi-account churn cascade risks.

• Model Trust & Algorithmic Fairness: Automated demographic parity and equalized odds testing ensuring strict adherence to the EEOC 4/5ths Rule across gender and sovereign territories.

• Survival Analysis & Dynamic CLTV: Kaplan-Meier survival curves and Cox Proportional Hazards hazard ratio modeling.

• Continuous MLOps (PSI Sentinel): Real-time Population Stability Index drift tracking between training baselines and live inference batches.

• AI Banking Copilot: Conversational quantitative assistant capable of answering complex portfolio queries with automated tool invocation.

How we built it

We engineered BankSight AI as a cohesive, production-grade enterprise stack:

• Machine Learning Core: Python 3.11, Scikit-learn, XGBoost, CatBoost, LightGBM, and SHAP (TreeExplainer). • High-Performance Backend: FastAPI with asynchronous lifespan management, typed Pydantic schemas, automated OpenAPI documentation (/docs), and native WebSocket event streaming. • Frontend Architecture: React 19, TypeScript, and Tailwind CSS, optimized with Vite 6. • 3D Graphics Engine: Three.js WebGL with custom raycasting, instanced meshes, and aspect-ratio-locked gyroscopic orbital animations. • Data Persistence & Audit: SQLite with ACID transaction safety and 10-point automated data quality validation. • DevOps & Cloud: Containerized with Docker, fully deployed on Render cloud infrastructure with live continuous integration.

Challenges we ran into Formulating the EARA Optimization Objective: High-risk customers often have small deposits, while affluent depositors often exhibit low baseline churn probabilities. Designing the EARA objective function required balancing capital exposure against model disagreement penalties to prevent costly false alarms. Causal Uplift Without Treatment Contamination: Separating true treatment uplift from organic retention without observational confounding required implementing dual XGBoost regressors and calculating synthetic Qini curves. WebGL 60 FPS Performance on Massive Point-Clouds: Rendering 10,004 individual 3D customer coordinates alongside real-time mouse parallax, dual rotating holographic orbs, and dynamic cluster filters required instanced mesh rendering and GPU offloading. Ensuring Regulatory Algorithmic Parity: Achieving a balanced 0.842 Disparate Impact ratio across protected demographic cohorts required calibrating classification thresholds rather than naive feature stripping. Accomplishments that we're proud of

• State-of-the-Art ML Performance: Reached 0.8612 ROC-AUC and an isotonic 0.1042 Brier Score on out-of-sample European retail test data. • Proprietary EARA Retention Engine: Successfully demonstrated that EARA retains €41.2M more capital under identical review budgets compared to standard greed-based churn rankings. • Flawless Enterprise Architecture: Delivered 22 distinct analytical modules with zero layout shift, seamless dark cyber-FinTech styling, and complete keyboard shortcuts (Ctrl + K Command Palette). • Production Cloud Deployment: 100% operational on cloud infrastructure (banksight-ai-o4ou.onrender.com) with live REST API documentation.

What we learned

• The Value of Explainable AI in Regulated Banking: TreeSHAP values are not just interpretability diagnostics—they can be transformed into actionable, customer-specific counterfactual recourse plans. • Causal Inference vs. Pure Prediction: Predicting who will churn is fundamentally different from predicting who can be saved. Causal ML protects banks from wasting outreach budgets on "Sleeping Dogs". • Enterprise FinTech UX: How to translate complex statistical formulations into intuitive, beautiful 3D dashboards that executives and quantitative analysts love using.

What's next for BankSight AI

• Direct Core Banking (CBS) Integration: Connecting live event listeners via SEPA / SWIFT messaging pipelines. • ECB Macro Stress-Testing: Extending the scenario simulator to simulate multi-quarter interest-rate shocks and systemic sovereign liquidity crunches. • Mobile Executive Terminal: Native iOS/Android companion app for on-the-go portfolio alerts and retention approval workflows.

Built With

• catboost
• docker
• fastapi
• lightgbm
• pandas
• python
• react
• render
• scikit-learn
• shap
• sqlite
• tailwindcss
• three.js
• typescript
• vite
• webgl
• websockets
• xgboost