Transaction Risk ML – Real-Time Risk Scoring System

📌 Overview

This project implements a real-time transaction risk scoring system using machine learning.

It predicts the fraud risk probability of a transaction and maps it into actionable decisions:

• ALLOW → Low risk
• CHALLENGE → Medium risk
• BLOCK → High risk

The system follows a production-style ML workflow with clearly separated training and inference pipelines.

---

Key Features

• End-to-end ML pipeline (training → inference)
• Real-time inference using FastAPI
• CI/CD integration with automated testing
• Dockerized deployment
• Structured logging (captured via cloud platform)
• Scenario-based validation for risk decisions

---

Model Summary

• Model: Logistic Regression
• Preprocessing: StandardScaler
• Training Data: Synthetic dataset with realistic fraud patterns
• ROC-AUC: ~0.93

The model outputs a risk probability, which is converted into business decisions using thresholding.

---

Project Structure

This project follows modular MLOps-inspired structure seprating training, inferenece, and deployment components.

Transaction-Risk-Scoring-System/
├── data/
├── training/
├── inference/
├── models/
├── utils/
├── tests/
├── .github/
│   └── workflows/
|
├── requirements.txt
├── Dockerfile
├── system_design.md
└── README.md

---

Running the Project

1. Install dependencies

pip install -r requirements.txt

2. Train model

python training/train.py

3. Run inference API

uvicorn inference.app:app --reload

API:

http://127.0.0.1:8000/docs

---

🌐 Live API

The model is deployed and accessible via a public API.

• Base URL https://transaction-risk-ml-production-system.onrender.com

• Swagger UI https://transaction-risk-ml-production-system.onrender.com/docs

---

Example Request

{ "amount": 2500, "account_age_days": 180, "past_txn_count_24h": 5, "hour_of_day": 22, "merchant_risk_score": 0.5 }

---

Example Response

{ "risk_score": 0.47, "decision": "CHALLENGE" }

---

⚡ Performance & Load Testing

The system was evaluated under simulated load to validate real-time performance.

• Tool Used: Locust
• Throughput: ~160 requests/second
• Latency: <150 ms (average)
• Failure Rate: 0% under sustained load

These results demonstrate the system’s ability to handle high-throughput, low-latency inference workloads.

---

Deployment

docker build -t /transaction-risk-ml . docker run -p 10000:10000 /transaction-risk-ml

The API is deployed on Render with logs streamed via stdout and monitored through the platform dashboard.

---

Dataset

The dataset used for training is synthetic and not included in this repository.

It simulates real-world fraud scenarios using:

• Transaction behavior
• Account characteristics
• Merchant risk signals

---

Model Performance

• Model: Logistic Regression
• ROC-AUC Score: ~0.93
• Pipeline: StandardScaler + Logistic Regression

---

Documentation

For detailed architecture, design decisions, and trade-offs:

➡️ Detailed System Design
→ Open

---