How we made our optical character recognition (OCR) code more accurate