Challenge
A large-scale consumer campaign required validating purchase receipts and verifying campaign participation across a very high volume of submissions. Manual verification was impossible at that scale, and accuracy directly affected who qualified for the campaign.
Approach
We built a validation pipeline combining AWS Textract for primary OCR extraction with Google's Gemini model as a reconciliation and validation layer. Receipts are read, key fields extracted, and participation criteria checked automatically. The two-model ensemble - dedicated OCR plus an LLM reconciliation pass - improves extraction reliability over either approach alone, and flags ambiguous cases rather than silently passing them.
Outcome
The campaign could validate receipts and participation automatically at a scale that manual review could never reach, with an ensemble design built to catch and surface edge cases instead of mis-accepting them.