Securing the Social Graph: An LLM Agent for Privacy-Preserving Analytics

As Large Language Models (LLMs) become more capable of analyzing complex personal datasets, users are finding ways to integrate their personal digital footprints into natural language interfaces. However, passing highly sensitive personal data to commercial, remote LLMs such as OpenAI or Gemini poses serious privacy risks to users' Personally Identifiable Information (PII). This project presents PARROT, an agentic workflow designed to safely preserve privacy on social media in the AsMoment app. The proposed architecture employs a Graph Database to model user chats and stories, enabling advanced Graph Retrieval-Augmented Generation (Graph RAG). A local LLM running on Llama 3 acts as the initial routing agent, preventing data leakage and generating database cipher queries to retrieve relevant context. To ensure the remote LLMs don't see users' data, Microsoft Presidio is integrated as a privacy layer. This layer intercepts and redacts PII from the database before it reaches the remote LLM for final synthesis. This report explains the system architecture, the prompt engineering required to force the local LLM to accurately generate Cypher queries, and the purpose of agentic redaction loops. The results show a functional, zero-leak pipeline that allows users to leverage the advanced reasoning of frontier models without exposing their raw personal data to third-party APIs. 

Major Professor: Simanta Mitra