Curious about the stack. Are you using Tesseract, or did you build your own OCR?

Also, it might start to get REALLY expensive to run this once you start getting any appreciable number of queries.