Identifying Synthetically Generated Documents and Receipts for Fraud Detection

This session will demonstrate an agentic fraud-screening pipeline that fuses computer vision with OCR-plus-LLM reasoning and operationalizes both in SAS Intelligent Decisioning. Documents and images are parsed via OCR; a language model evaluates numerical and semantic consistency to flag anomalies characteristic of synthetic artifacts. In parallel, a computer vision model projects inputs into an embedding space and uses distance-based signals to detect image generation artifacts. A global decision model then combines the CV distance features with the OCR/LLM outputs to produce a calibrated risk score that supports policy-driven actions (auto-approve, human review, reject). The solution is orchestrated end-to-end in SAS Viya with execution flows and dashboards for monitoring and extension to new modalities.

About the Presenter

Robert is a Principal Data Scientist at SAS where he builds end-to-end artificial intelligence applications.  He also researches, consults, and teaches machine learning with an emphasis on deep learning and computer vision for SAS. Robert has authored an introductory book on computer vision and has written several professional courses on topics including neural networks, deep learning, and optimization modeling.  
 
Before joining SAS, Robert worked under the Senior Vice Provost at North Carolina State University, where he built models pertaining to student success, faculty development, and resource management. Prior to working in academia, Robert was a member of the research and development group on the Workforce Optimization team at Travelers Insurance. His models at Travelers focused on forecasting and optimizing resources.
 
Robert graduated with a master’s degree in Business Analytics and Project Management from the University of Connecticut and a master’s degree in Applied and Resource Economics from East Carolina University.