🛡️ FraudShield: Enterprise End-to-End ML Platform

A production-grade, self-healing machine learning platform for Real-time Credit Card Fraud Detection. This platform integrates the entire ML lifecycle—from experiment tracking and model governance to autonomous retraining and explainable AI.

🏗️ System Architecture

graph TD
    subgraph "External Layer"
        User[Investigator / Client]
    end

    subgraph "Serving Layer (FastAPI + Gunicorn)"
        API[Secure API Endpoint]
        Val[Data Validation - Pydantic]
        API --> Val
    end

    subgraph "Intelligence Layer"
        Champion[Champion Model - RF]
        SHAP[Explainability Engine - SHAP]
        Val --> Champion
        Champion --> SHAP
    end

    subgraph "Governance & Storage"
        MLflow[MLflow Model Registry]
        Feast[Feature Store]
        Champion --- MLflow
        Champion --- Feast
    end

    subgraph "Monitoring & Maintenance"
        Monitor[Drift Monitor - Evidently]
        Heal[Self-Healing Orchestrator]
        Monitor --> Heal
        Heal --> MLflow
    end

    subgraph "UI Layer"
        Dash[Streamlit Cockpit]
        PDF[PDF Report Generator]
        Dash --> API
        Dash --> PDF
    end

    User --> API
    User --> Dash
    Monitor --- API

Loading

🚀 Key Features

• 🧠 Explainable AI (SHAP): Every prediction comes with a detailed transparency report, showing exactly which features (amount, time, category) influenced the risk score.
• 🏆 Model Registry (Champion/Challenger): Professional governance using MLflow to manage model versions and aliases. The API automatically serves the "Champion" model.
• 🔄 Self-Healing Pipeline: An autonomous orchestrator that detects data drift and automatically triggers retraining and model promotion.
• 📊 Management Cockpit: A beautiful Streamlit dashboard for fraud investigators to run manual checks, view SHAP visualizations, and monitor system health.
• 📄 Professional PDF Reports: Generate and download comprehensive analysis reports for any transaction, featuring embedded SHAP charts, system flow diagrams, and audit details.
• 📜 Dynamic Audit Logs: Real-time session-based history of all analyzed transactions, providing a traceable path for manual investigations.
• 🛡️ Multi-Layer Validation: Incoming data is strictly validated via Pydantic and Great Expectations before reaching the model.
• ⚡ Production Optimized: Containerized with Gunicorn and multi-worker Uvicorn for high-throughput and low-latency serving (<200ms).

🛠️ Technology Stack

Component	Technology
Model Serving	FastAPI, Gunicorn, Uvicorn
ML Lifecycle	MLflow (Tracking & Registry)
Explainability	SHAP (Shapley Additive Explanations)
Monitoring	Evidently AI (Drift & Data Quality)
Management UI	Streamlit
Data Validation	Pydantic, Great Expectations
Feature Store	Feast
Orchestration	Python-based Autonomous Orchestrator
Infrastructure	Docker, Docker Compose
ML Core	Scikit-learn (Random Forest)

📁 Project Structure

END_TO_END_ML_PLATFORM/
├── src/
│   ├── serve_validated.py    # Secured, Explainable API
│   ├── train_advanced.py     # MLflow-integrated training
│   ├── orchestrator.py       # Self-healing autonomous loop
│   ├── dashboard.py          # Streamlit Management Cockpit
│   ├── promote_model.py      # Automated Model Promotion
│   ├── data_validation.py    # Input validation rules
│   └── monitoring.py         # Drift detection logic
├── feature_repo/             # Feast Feature Store definitions
├── models/                   # Local model artifacts
├── data/                     # Training and Production datasets
├── Dockerfile                # Multi-worker production build
├── docker-compose.yml        # Full-stack orchestration
└── mlflow.db                 # Model Registry database

⚡ Quick Start (Production Mode)

The entire platform can be launched as a unified stack using Docker Compose:

# 1. Clone the repository
git clone https://github.com/LEVELING2108/End-to-End-ML-Platform---Fraud-Detection.git
cd End-to-End-ML-Platform---Fraud-Detection

# 2. Launch the full stack
docker-compose up --build

Access Points:

• Fraud API: http://localhost:8000/docs (Requires X-API-KEY)
• Management Dashboard: http://localhost:8501
• MLflow Tracker: http://localhost:5000

🔒 API Security

The prediction endpoints are secured. To interact with the API, include the following header: X-API-KEY: fraud-detection-secret-key

📈 Performance Benchmarks

• Throughput: ~50+ transactions per second (Horizontal scalable)
• Latency: Average ~120ms per prediction (v4.0.0 with Gunicorn workers)
• Stability: 100% success rate under 50 concurrent user load.

---

📚 Acknowledgments & References

This project was originally inspired by and based on the FreeCodeCamp End-to-End ML Platform tutorial. It has since been expanded into a commercial-grade platform with the addition of SHAP explainability, automated model promotion, and self-healing orchestration.