Fraud Detection Inference Service

Real-time fraud detection inference service.
Built with Celery + RabbitMQ as message broker, results and transaction data stored in PostgreSQL.
Uses XGBoost model with strong class imbalance handling via SMOTE.
Hyperparameters were optimized using Optuna.

Architecture

• Backend: Celery workers
• Message Broker: RabbitMQ
• Database: PostgreSQL
• ML Model: XGBoost (trained with SMOTE + Optuna tuning)
• Deployment: Docker + Docker Compose

Main Components

• app/app.py — Celery configuration and task registration
• worker/tasks.py — main task run_fraud_inference(**data_dict) for model inference
• models/ — serialized XGBoost model (xgb_final.joblib) + preprocessing pipeline
• docker-compose.yml — launches RabbitMQ, PostgreSQL and Celery workers

Quick Start (Local)

Requirements

• Docker
• Docker Compose

Run

docker-compose up --build

Default flower localhost:5555