SentinelFlow: Machine Learning Engineering Practice for Production-Ready Financial Fraud Detection

SentinelFlow is an end-to-end financial fraud detection platform for production environments, with three core modules: traditional machine learning pipeline, real-time inference service, and graph neural network relationship analysis. The project is designed with a "production-first" philosophy, demonstrating how to smoothly transition fraud detection models from experimental environments to production systems, making it a highly valuable open-source learning resource in the fintech field.

In the financial industry, fraud detection is one of the most mature yet challenging areas for machine learning applications. Traditional rule engines are stable but struggle to handle complex fraud methods; purely research-oriented ML models lack an engineering implementation path. SentinelFlow was born to address this pain point—it is not just an algorithm prototype, but a complete production-grade ML engineering practice that follows industrial standards, providing open-source learning resources for fintech engineers and researchers.

SentinelFlow adopts a phased architecture:

1. Traditional ML pipeline (baseline built with Scikit-learn + XGBoost)
2. Production-grade real-time inference platform (real-time services provided by FastAPI)
3. Graph neural network relationship analysis (network fraud identification implemented with PyTorch Geometric) In terms of tech stack selection, Python is the main language, combined with PostgreSQL for persistence and Docker for standardized deployment, forming a lightweight and complete technical closed loop.

Using the Kaggle Credit Card Fraud Dataset (2013 European transaction records, fraud ratio 0.172%), feature engineering uses a combination of standardized, PCA-reduced anonymous features and original features. Models compared include logistic regression, random forest, and XGBoost—XGBoost performs best due to its ability to handle non-linear relationships and regularization. Evaluation uses metrics such as precision-recall curve, F1 score, and AUC-ROC, balancing the business trade-offs between false negatives and false positives.

A RESTful service is built with FastAPI, using asynchronous features to handle concurrency and automatically generating OpenAPI documentation to reduce joint debugging costs. Model version management and hot loading are implemented—after a new model is validated, it can be switched seamlessly without restarting. It includes a monitoring and logging module to record request data, and PostgreSQL stores transaction and prediction history to support auditing. Docker containerization ensures environment consistency, enabling rapid deployment to local or cloud servers.

Traditional models ignore transaction correlations. The GNN module uses PyTorch Geometric to model accounts/transactions as nodes/edges—node features include account attributes/behaviors, and edge features encode transaction amount/time. Through GCN/GAT message passing, multi-hop neighbor information is aggregated to identify gang fraud. It complements traditional models: traditional models handle real-time single-transaction scoring, while GNN is used for offline in-depth analysis, building a layered defense system.

The project demonstrates the end-to-end process of ML system implementation (data → features → training → deployment → monitoring → iteration), which is of great significance for developers transitioning from academia to industry. The tech stack covers traditional ML (Scikit-learn/XGBoost), cloud-native development (FastAPI/Docker), and cutting-edge DL (PyTorch Geometric), forming an evolution path from traditional to modern. As a baseline framework, it can integrate real data and enterprise-level tools (MLflow/Prometheus) to build customized platforms.

SentinelFlow is a well-structured ML engineering project that demonstrates the evolution path of fraud detection technology. Future expansion directions:

1. Online learning to achieve adaptive model updates
2. Integrate device fingerprints/behavioral biometrics to enrich features
3. Federated learning to share fraud information across institutions
4. Explainable AI to enhance decision transparency and meet compliance requirements