The latest articles on DEV Community by Bhagyashri Kale (@bhagyashri_kale_c500cb884). https://dev.to/bhagyashri_kale_c500cb884 https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4142189%2Fc219cbe4-a091-4436-9588-132d3e908fc6.png https://dev.to/bhagyashri_kale_c500cb884 en Bhagyashri Kale Fri, 25 Sep 2026 03:52:58 +0000 https://dev.to/bhagyashri_kale_c500cb884/agentic-fraud-investigation-with-tigergraph-2lmi https://dev.to/bhagyashri_kale_c500cb884/agentic-fraud-investigation-with-tigergraph-2lmi <h1> Agentic Fraud Investigation with TigerGraph </h1> <h2> Introduction </h2> <p>Fraud investigation becomes difficult when transaction data, customer information, and related entities are spread across different records.</p> <p>For the TigerGraph × HHGoa 2026 challenge, we built an Agentic Fraud Investigation solution to analyze fraud cases and generate structured investigation results.</p> <p>Our solution uses TigerGraph Savanna as the graph layer and a custom Python-based investigation agent.</p> <h2> Problem Statement </h2> <p>The objective was to investigate fraud cases using transaction and entity relationships and produce useful evidence and investigation decisions for each case.</p> <p>Our solution processes 20 benchmark cases, from HHG-001 to HHG-020.</p> <h2> What We Built </h2> <p>Our solution includes:</p> <ul> <li>TigerGraph Savanna as the graph layer</li> <li>SavannaTransactionGraph</li> <li>Transaction, Customer, and Card entities</li> <li>Relationships between connected entities</li> <li>A custom Python investigation agent</li> <li>Rule-based fraud pattern detection</li> <li>Heuristic risk scoring</li> <li>Structured JSON outputs for each case</li> </ul> <h2> Architecture </h2> <p>The workflow is:</p> <p>Case Input<br><br> → Data Preparation<br><br> → TigerGraph Graph<br><br> → Transaction and Entity Relationships<br><br> → Python Investigation Agent<br><br> → Fraud Pattern Analysis<br><br> → Risk Score<br><br> → Investigation Decision<br><br> → JSON Case Output</p> <p>The graph representation helps organize relationships between transactions and related entities for fraud investigation.</p> <h2> How We Used TigerGraph </h2> <p>We used TigerGraph Savanna as the graph layer for our fraud investigation workflow.</p> <p>The graph used in the project is:</p> <p><code>SavannaTransactionGraph</code></p> <p>The graph contains entities such as:</p> <ul> <li>Transaction</li> <li>Customer</li> <li>Card</li> </ul> <p>These entities are connected through relationships that help organize transaction-related evidence.</p> <h2> Investigation Agent </h2> <p>The investigation agent is implemented as a custom Python program.</p> <p>It reads the prepared investigation data, applies fraud detection rules, calculates a heuristic risk probability, and generates investigation decisions.</p> <p>The current implementation is rule-based and heuristic. It does not use an external LLM API.</p> <h2> Results </h2> <p>We successfully processed all 20 benchmark cases:</p> <p><code>HHG-001</code> to <code>HHG-020</code></p> <p>The final results are stored as individual JSON case files.</p> <p>These outputs contain structured investigation information for each case, including the detected patterns and investigation decision.</p> <h2> GitHub Repository </h2> <p>The complete project source code and case outputs are available here:</p> <p><a href="https://github.com/Bhagyashrikale01/agentic-fraud-investigation-tigergraph" rel="noopener noreferrer">https://github.com/Bhagyashrikale01/agentic-fraud-investigation-tigergraph</a></p> <h2> Limitations </h2> <p>The current implementation has some limitations:</p> <ul> <li>The risk probability is heuristic and is not a trained or validated probability.</li> <li>The investigation uses preprocessed evidence.</li> <li>The current local agent does not use live external integrations.</li> <li>Full direct TigerGraph MCP integration with the agent is planned as future work.</li> </ul> <h2> Future Improvements </h2> <p>Future improvements could include:</p> <ul> <li>Full TigerGraph MCP integration</li> <li>GraphRAG-based investigation</li> <li>LLM-assisted investigation reasoning</li> <li>Live data integrations</li> <li>More advanced fraud detection models</li> <li>Improved explainability and investigation reports</li> </ul> <h2> Conclusion </h2> <p>This project gave us practical experience in graph-based fraud investigation, transaction analysis, Python automation, and building an end-to-end investigation workflow.</p> <p>The TigerGraph × HHGoa 2026 challenge helped us understand how graph technology can connect related entities and support fraud investigation.</p> <p>Thank you to TigerGraph and the HHGoa team for the opportunity to work on this challenge.</p> <h1> TigerGraph #FraudDetection #Python #Hackathon </h1> ai database devchallenge python Bhagyashri Kale Fri, 25 Sep 2026 03:46:51 +0000 https://dev.to/bhagyashri_kale_c500cb884/agentic-fraud-investigation-with-tiger-graph-3mk5 https://dev.to/bhagyashri_kale_c500cb884/agentic-fraud-investigation-with-tiger-graph-3mk5 <h1> Agentic Fraud Investigation with TigerGraph </h1> <h2> Introduction </h2> <p>Fraud investigation becomes difficult when transaction data, customer information, and related entities are spread across different records.</p> <p>For the TigerGraph × HHGoa 2026 challenge, we built an Agentic Fraud Investigation solution to analyze fraud cases and generate structured investigation results.</p> <p>Our solution uses TigerGraph Savanna as the graph layer and a custom Python-based investigation agent.</p> <h2> Problem Statement </h2> <p>The objective was to investigate fraud cases using transaction and entity relationships and produce useful evidence and investigation decisions for each case.</p> <p>Our solution processes 20 benchmark cases, from HHG-001 to HHG-020.</p> <h2> What We Built </h2> <p>Our solution includes:</p> <ul> <li>TigerGraph Savanna as the graph layer</li> <li>SavannaTransactionGraph</li> <li>Transaction, Customer, and Card entities</li> <li>Relationships between connected entities</li> <li>A custom Python investigation agent</li> <li>Rule-based fraud pattern detection</li> <li>Heuristic risk scoring</li> <li>Structured JSON outputs for each case</li> </ul> <h2> Architecture </h2> <p>The workflow is:</p> <p>Case Input<br><br> → Data Preparation<br><br> → TigerGraph Graph<br><br> → Transaction and Entity Relationships<br><br> → Python Investigation Agent<br><br> → Fraud Pattern Analysis<br><br> → Risk Score<br><br> → Investigation Decision<br><br> → JSON Case Output</p> <p>The graph representation helps organize relationships between transactions and related entities for fraud investigation.</p> <h2> How We Used TigerGraph </h2> <p>We used TigerGraph Savanna as the graph layer for our fraud investigation workflow.</p> <p>The graph used in the project is:</p> <p><code>SavannaTransactionGraph</code></p> <p>The graph contains entities such as:</p> <ul> <li>Transaction</li> <li>Customer</li> <li>Card</li> </ul> <p>These entities are connected through relationships that help organize transaction-related evidence.</p> <h2> Investigation Agent </h2> <p>The investigation agent is implemented as a custom Python program.</p> <p>It reads the prepared investigation data, applies fraud detection rules, calculates a heuristic risk probability, and generates investigation decisions.</p> <p>The current implementation is rule-based and heuristic. It does not use an external LLM API.</p> <h2> Results </h2> <p>We successfully processed all 20 benchmark cases:</p> <p><code>HHG-001</code> to <code>HHG-020</code></p> <p>The final results are stored as individual JSON case files.</p> <p>These outputs contain structured investigation information for each case, including the detected patterns and investigation decision.</p> <h2> GitHub Repository </h2> <p>The complete project source code and case outputs are available here:</p> <p><a href="https://github.com/Bhagyashrikale01/agentic-fraud-investigation-tigergraph" rel="noopener noreferrer">https://github.com/Bhagyashrikale01/agentic-fraud-investigation-tigergraph</a></p> <h2> Limitations </h2> <p>The current implementation has some limitations:</p> <ul> <li>The risk probability is heuristic and is not a trained or validated probability.</li> <li>The investigation uses preprocessed evidence.</li> <li>The current local agent does not use live external integrations.</li> <li>Full direct TigerGraph MCP integration with the agent is planned as future work.</li> </ul> <h2> Future Improvements </h2> <p>Future improvements could include:</p> <ul> <li>Full TigerGraph MCP integration</li> <li>GraphRAG-based investigation</li> <li>LLM-assisted investigation reasoning</li> <li>Live data integrations</li> <li>More advanced fraud detection models</li> <li>Improved explainability and investigation reports</li> </ul> <h2> Conclusion </h2> <p>This project gave us practical experience in graph-based fraud investigation, transaction analysis, Python automation, and building an end-to-end investigation workflow.</p> <p>The TigerGraph × HHGoa 2026 challenge helped us understand how graph technology can connect related entities and support fraud investigation.</p> <p>Thank you to TigerGraph and the HHGoa team for the opportunity to work on this challenge.</p> <h1> TigerGraph #FraudDetection #Python #Hackathon </h1>