Building a File Semantic Analyzer: Guarding Outbound Data at Scale with AI

This post details the engineering journey of Uber's File Semantic Analyzer (FSA), an AI-powered system for classifying outbound data. It elaborates on the problem of traditional DLP systems' limitations, the solution leveraging Generative AI for semantic understanding, and the engineering steps involved: data labeling, pre-processing (content extraction, OCR, intelligent chunking), the GenAI interpretation engine (summarization, entity extraction, intent inference, explainability), and the classification/policy enforcement layer. The architecture is described, highlighting the File Connector, File Processing, Prompt Building, Decision engine, and Manual Review components. Risks such as LLM hallucinations and context loss are addressed through explanations, intelligent chunking, and human-in-the-loop validation. The impact includes accelerated incident response times and significant time savings for security analysts.