PAR Tech builds multi-tenant LLM analytics with row-level security on AWS

PAR Technology Corporation, a restaurant tech provider serving over 300 businesses from independent operators to large franchise groups, built a multi-tenant LLM analytics system that solves the critical challenge of row-level security. The system uses a natural language text-to-SQL agent where business users can ask questions in plain English and get answers in seconds. However, the same question from different users must return different results based on their data permissions—a franchise owner sees only their two stores' sales ($84,000), while a brand manager sees national totals ($9.2 million).

To enforce this deterministically, PAR implemented a three-layer architecture: (1) cryptographic request signing with AWS SigV4 to authenticate users, (2) semantic validation on Amazon Bedrock to verify query intent, and (3) programmatic data isolation via Split-Plane SQL that injects tenant-specific filters at the database level. This design prevents cross-tenant data exposure even if the LLM is compromised or hallucinates, because security is enforced at the architecture level, not by trusting the model's behavior. The system handles thousands of queries daily while maintaining strict data boundaries.