OCR as a use case for LLMs cracks me up because traditional NN OCR is probably more accurate, but significantly more efficient. I figured people would chain it with LLMs to fix misscans.