Detect payroll discrepancies without sending pay stubs offsite
AI Pay Stub Compliance Product
Payment discrepancy detection on a self-hosted Courier stack
The situation
The problem
A founder was building a payments tracking product that used AI to discover complex payment discrepancies. The inputs were pay stubs and payroll records — data you do not casually upload to a public model.
How it usually works
The default is offsite OCR and a hosted LLM. Stubs, names, amounts, and employer details leave the product, sit in a provider's pipeline, and may be retained for abuse review or training. Privacy policy becomes a PDF, not an architecture.
How we ran it
Privacy requirements meant self-hosted AI. We ran Gemma 3 27B on Courier so analyses, extraction, and discrepancy discovery happen on infrastructure the client controls. Pay data does not need to leave the perimeter to get a finding.
Strategy
Treat payroll documents as a perimeter problem first and a model problem second. The feature has to be affordable and performant without a third party seeing the stubs.
Implementation
Self-hosted Gemma 3 27B on Courier for structured analyses, data extraction, and discrepancy discovery with clear guidelines — vLLM, OpenAI-compatible, on a stack they operate.
Results
- A production AI feature that was affordable, performant, and ran on infrastructure the client controlled.