Enterprise Fraud Detection (EFD)

A comprehensive enterprise fraud detection system built on Palantir Foundry, featuring real-time transaction processing, machine learning-based fraud prediction, and an interactive web dashboard.

Architecture Overview

Kafka Stream → Transaction Processing → Feature Engineering → ML Inference → Ontology → Dashboard

Components

Component	Directory	Description
Data Pipeline	python-transforms/	Streaming transaction ingestion, Avro-to-Parquet conversion, and ML inference transforms
ML Model	model-training/	XGBoost + Random Forest ensemble model with threshold optimization
Training Data	training-data-generation/	Synthetic customer and transaction data generator with fraud labeling
Web Dashboard	efd-website/	React OSDK application for fraud monitoring and investigation

---

Data Pipeline (python-transforms/)

The data pipeline handles real-time transaction ingestion and fraud scoring:

• transform.py — Streaming transaction processor that ingests from Kafka, deduplicates, and enriches transactions
• fraud_predictions.py — Applies the trained ML model to score transactions in real-time
• avro_to_parquet.py — Converts raw Avro-encoded Kafka messages to structured Parquet format

Key Features

• Incremental processing (only new transactions since last build)
• Automatic schema evolution handling
• Configurable fraud threshold scoring

---

ML Model (model-training/)

The fraud detection model uses an ensemble approach combining XGBoost and Random Forest classifiers.

Training Pipeline

• feature_engineering.py — Feature extraction and transformation (transaction velocity, amount statistics, time-based features)
• model_training.py — Model training with hyperparameter tuning and threshold optimization for precision/recall trade-off
• generated_adapter.py — FraudDetectionModelAdapter for Foundry model deployment
• run_inference.py — Batch inference pipeline
• adapters.py — Model adapter interfaces

Model Performance

• Optimized decision threshold balancing false positives vs. missed fraud
• Feature importance tracking for model interpretability
• Version-controlled model artifacts with full lineage

---

Training Data Generation (training-data-generation/)

• examples.py — Generates synthetic customer profiles and transaction histories with realistic fraud patterns
  • Normal spending behavior simulation
  • Fraud injection with configurable fraud rate
  • Customer demographic generation
  • Temporal patterns (time-of-day, day-of-week effects)

---

Web Dashboard (efd-website/)

A React application built with the Palantir Ontology SDK (OSDK) providing:

• Real-time fraud alerts — Live monitoring of flagged transactions
• Investigation workflows — Drill-down into suspicious activity patterns
• Customer profiles — 360° view of customer behavior and risk scores
• Analytics dashboards — Aggregate fraud metrics and trends

Tech Stack

• React + TypeScript
• Vite build system
• Palantir OSDK for Ontology integration
• Responsive design for desktop and mobile

---

Getting Started

This project is built and run on Palantir Foundry. The source code is exported here for reference and version control.

Prerequisites

• Palantir Foundry enrollment with:
  • Data Connection (Kafka connector)
  • Transforms (Python)
  • Model Training
  • OSDK application hosting

Project Structure

enterprise-fraud-detection-model/
├── README.md
├── python-transforms/       # Data pipeline transforms
│   ├── pipeline.py
│   ├── datasets/
│   │   ├── transform.py
│   │   ├── fraud_predictions.py
│   │   └── avro_to_parquet.py
│   └── ...
├── model-training/          # ML model code
│   ├── model_training.py
│   ├── feature_engineering.py
│   ├── generated_adapter.py
│   ├── run_inference.py
│   └── ...
├── training-data-generation/ # Synthetic data generation
│   └── examples.py
└── efd-website/             # React OSDK dashboard
    ├── package.json
    ├── src/
    │   ├── main.tsx
    │   ├── client.ts
    │   └── ...
    └── ...

---

License

Proprietary. All rights reserved.

---

Export Details

This repository is automatically synced from Palantir Foundry using a custom export pipeline. Changes should be made in Foundry and re-exported.

Last export method: Foundry → GitHub Contents API (automated pipeline)