Theia Squad: A Real-Time Transaction Fraud Detection System Based on Multi-Agent Architecture

Theia Squad is a modular multi-agent system that combines vector similarity retrieval, machine learning classification, and AI reasoning to achieve real-time transaction fraud detection and interpretable decision-making. Its core lies in a three-agent collaborative architecture, which not only addresses the limitations of traditional methods but also meets the interpretability requirements for regulatory compliance.

With the popularity of digital payments, financial fraud has become a major challenge. Traditional rule engines and single models struggle to cope with evolving fraud methods and lack interpretability. Theia Squad adopts an innovative multi-agent architecture, combining similarity retrieval, machine learning classification, and AI reasoning to build a real-time detection and interpretable system.

Theia Squad decomposes the task into three specialized agents:

1. Retrieval Agent: Uses cosine similarity to search historical transactions and obtain the top 3 similar fraud cases as context;
2. Classification Agent: Uses a logistic regression model to predict fraud probability, enhancing results with similar patterns;
3. Reasoning Agent: Integrates classification results and context to generate interpretable decisions and confidence levels. The process is Transaction → Retrieval → Classification → Reasoning.

• Vector Storage: In-memory vector storage + PCA dimensionality reduction to visualize fraud pattern distribution;
• Machine Learning Model: Default logistic regression with training parameters: 5 epochs, batch size 1024, learning rate 0.001, supporting processing of 5 million transactions;
• Explainable AI: Uses similar case references + natural language reasoning explanations, such as specific feature descriptions (e.g., high-value transfers).

The project offers two usage methods:

• Command Line: Suitable for batch processing;
• Streamlit Web Dashboard: Real-time interaction, including transaction selection, fraud probability visualization, reasoning explanations, PCA embedding charts, and similar pattern displays. The startup command is concise, with a local service on port 8502.

Current Capabilities: Demonstrates the potential of multi-agents in financial risk control; modular design facilitates optimization and replacement; similarity retrieval provides traceable basis; meets compliance requirements; Known Limitations: Uses synthetic data by default; in-memory vector storage has scalability bottlenecks; logistic regression model is simple (for demonstration purposes, needs upgrade for production).

• Data: Integrate CIS-IEEE anonymous dataset; replace in-memory vector storage with FAISS;
• Model: Introduce LangChain/LangGraph framework to enhance large language model reasoning;
• Deployment: AWS SageMaker cloud deployment; integrate Kafka real-time stream processing; Ray distributed reasoning;
• Operations: Add model monitoring and drift detection; establish an A/B testing framework.

Theia Squad provides an architectural reference for financial fraud detection. Multi-agents address the limitations of single models; interpretability meets compliance requirements; modularity reserves space for evolution. Production-level systems need to select appropriate data sources, vector storage, and reasoning models, and establish monitoring and iteration mechanisms.