Trace Me If You Can

Even Escobar couldn’t escape this net.

Comment

Inspiration

Money laundering moves an estimated $800 billion to $2 trillion globally each year, yet most Anti-Money Laundering (AML) tools are still rule-based and flat. These systems analyze transactions individually and generate huge numbers of false positives while still missing sophisticated laundering networks.

In reality, money laundering is dynamic and heavily networked. Funds move across many accounts, institutions, and time intervals, making it difficult for investigators to detect patterns using traditional tools.

We built this project to rethink AML investigation using graph machine learning and generative AI, giving financial investigators a system that can analyze transaction networks, explain suspicious activity, and support real investigations.

Product Summary

Our platform is an AI-powered Anti-Money Laundering investigation system designed for financial investigators.

Instead of analyzing transactions in isolation, the system builds a transaction graph where:

  • Accounts are nodes
  • Transactions are edges

This allows investigators to visualize laundering activity as a network, revealing hidden patterns and suspicious relationships between accounts.

Core Workflow

  1. Transaction data from the IBM AML dataset is ingested into the system.
  2. The system constructs a transaction graph, where accounts are nodes and transactions are edges.
  3. A Graph Neural Network (GraphSAGE) analyzes the network and assigns a risk score to each account.
  4. The model intentionally over-flags accounts with suspicious signals to ensure potential laundering activity is not missed.
  5. Flagged accounts are analyzed by three AI agents orchestrated through Railtracks.
  6. Each agent evaluates the case independently and produces a structured JSON analysis describing suspicious behaviors and risk indicators.
  7. The combined JSON output is sent to IBM Watsonx, which synthesizes the agents’ findings and converts them into a clear, human-readable explanation.
  8. Investigators review the results through the interactive dashboard and network visualization tools.
  9. The system can generate structured investigation reports, allowing investigators to easily document findings and escalate suspicious cases for compliance review.

Key Features

Network Investigation

  • Interactive 2D / 3D transaction network
  • Risk-colored nodes showing suspicious accounts
  • Click any account to inspect its activity and connections

Graph Neural Network Detection

At the core of the system is a Graph Neural Network (GraphSAGE) used to detect suspicious accounts within the transaction network.

Instead of analyzing transactions individually, the model learns patterns directly from the structure of the financial network.

Graph Representation

  • Nodes: Bank accounts
  • Edges: Transactions between accounts
  • Graph Size: ~515k accounts and ~5 million transactions in the IBM AML dataset

Model Architecture

We implemented several GNN architectures (GCN, GraphSAGE, and GAT), with GraphSAGE selected as the production model because of its scalability and strong performance on large graphs.

Key configuration:

  • 3 GraphSAGE layers
  • 128 hidden dimensions
  • Max-pooling neighbor aggregation
  • Residual connections
  • Dropout regularization

Input Features

Each account node is represented by a 58-dimensional feature vector including:

  • Transaction statistics (counts, sums, averages)
  • Network structure signals
  • Temporal activity patterns
  • Payment format diversity
  • Random-walk features that capture circular money flows

These features allow the model to detect behaviors commonly associated with laundering such as:

  • Circular fund transfers
  • Hub accounts aggregating funds
  • Rapid movement of money across accounts

Model Output

The model produces a risk score between 0 and 1 for every account.

Accounts above a threshold are flagged and passed to the Railtracks multi-agent analysis system, where AI agents investigate and explain the suspicious activity.

This approach allows the system to detect network-level laundering behavior that rule-based systems often miss.

AI Investigation Panel

Each account includes:

  • Risk score
  • Transaction volume
  • Counterparties
  • Detected patterns
  • AI-generated explanation of suspicious behavior

Multi-Agent AI Analysis

Flagged accounts are analyzed by three AI agents orchestrated through Railtracks.

Each agent evaluates the case independently and produces structured analysis about the account's behavior, transaction patterns, and risk indicators. The agents review the case from different perspectives to reduce bias and improve reliability.

The results from the three agents are exported as structured JSON. This JSON output is then sent to IBM Watsonx, which synthesizes the agents’ findings and converts them into a clear, human-readable explanation for investigators.

This two-step system allows the platform to combine structured AI analysis with natural language reasoning, making complex machine learning outputs easier for financial investigators to understand.

Investigator Copilot

The AI acts as a copilot for financial investigators, helping them:

  • understand machine learning signals
  • interpret suspicious transaction patterns
  • quickly review flagged accounts

Investigation Reports

Investigators can generate clean summaries of flagged accounts for documentation and compliance reporting.

Technology Stack

Languages

  • Python
  • TypeScript
  • JavaScript

Frameworks and Libraries

  • FastAPI for backend APIs
  • Next.js + React for the frontend
  • PyTorch Geometric for Graph Neural Networks
  • Three.js + react-force-graph for graph visualization
  • Tailwind CSS for UI design

AI / Machine Learning

  • GraphSAGE Graph Neural Network
  • Railtracks multi-agent orchestration
  • Gemini 3.0 for AI reasoning and explanation

Platforms

  • IBM Cloud
  • IBM Watsonx
  • IBM Db2

Data

  • IBM Anti-Money Laundering Dataset (Kaggle)

Tools

  • Docker
  • SQLite
  • Pandas
  • NumPy

AI Use

Yes. Approximately 70% of the codebase was generated or assisted by AI tools during development.

AI is also a core component of the system:

  • Graph Neural Networks detect suspicious financial behavior within transaction networks.
  • Generative AI agents explain flagged accounts and assist investigators in understanding potential laundering activity.

GitHub Repository

Include the link to the public repository below:

https://github.com/phintruong/Trace-Me-If-You-Can

The repository contains:

  • Backend AML pipeline
  • Graph neural network inference
  • Multi-agent AI explanation system
  • Investigation dashboard
  • Setup instructions and documentation

Built With

Share this project:

Updates