I made a tool like that, and I bet with a more powerful LLM like GPT4, and perhaps a better baseline OCR tool (like GPT4 vision), it could work really well for this sort of thing:

https://github.com/Dicklesworthstone/llama2_aided_tesseract