FlowTrace.exe — AI Fund Flow Fraud Detection

• 

  Running Detection Pipeline

• 

  Risk Score Distribution for different account and transactions

• 

  Dashboard for Transaction Timeline

• 

  Generating Artificial Dataset

• 

  Transaction Network Mesh

Inspiration

Banks process millions of transactions daily, but fraud is rarely visible in a single transaction. Instead, it is hidden across multiple accounts, branches, and channels using techniques like layering, round-tripping, and structuring.

We were inspired by the limitation of current systems that fail to track the complete journey of money. This motivated us to build a system that detects, traces, and explains fraud effectively.

What it does

FlowTrace.exe is an AI-powered fund flow tracking system that:

• Tracks end-to-end fund movement
• Uses graph analytics to detect hidden relationships and fraud networks
• Applies machine learning for anomaly detection
• Generates LLM-powered explanations
• Produces investigator-ready reports

How we built it

We designed the system in multiple layers:

• Data Ingestion: Accepts CSV/Excel transaction data and normalizes it
  
• Graph Construction:
  Nodes = accounts/customers
  Edges = transactions
  
• Detection Engine:
  Combines rule-based detection, machine learning, and graph analytics
  
• Risk Scoring: Assigns risk scores to transactions and accounts
  
• LLM Layer: Generates explanations using APIs like Gemini/Claude/Mistral
  
• Dashboard: Visualizes fund flow and suspicious paths
  

Example code:

risk_score = abs(observed - expected) / expected

Inline math: ( Risk = \frac{|Observed - Expected|}{Expected} )

Block math: $$ Risk\ Score = \frac{|Observed - Expected|}{Expected} $$

Challenges we ran into

• Handling large-scale transaction data efficiently
• Reducing false positives
• Integrating ML, graph analytics, and LLMs
• Managing incomplete and inconsistent data
• Ensuring clear and explainable outputs

Accomplishments that we're proud of

• Built a graph-based fraud detection system
• Enabled multi-hop fund tracing
• Implemented a hybrid detection engine
• Integrated LLM-based explanations
• Generated automated evidence reports

What we learned

• Graph-based analysis is crucial for detecting hidden fraud
• Combining rules + ML + graph analytics improves accuracy
• Explainable AI is essential in financial systems
• Real-world data is often noisy and incomplete
• Modular design improves scalability

What's next for FlowTrace.exe — AI Fund Flow Fraud Detection

• Add real-time streaming support
• Improve ML models with continuous learning
• Integrate external risk data sources
• Enhance dashboard with advanced analytics
• Deploy as a scalable cloud-based solution

Built With

• analytics
• claude
• cytoscape.js
• d3.js
• docker
• fastapi
• gemini
• graph
• learning
• machine
• mistral
• networkx
• numpy
• pandas
• plotly
• postgresql
• python
• react
• scikit-learn
• sqlite
• streamlit