Painful comparison haha

Leaving a comment so I can more easily find this

And for the people wondering about Pandas, use Polars instead

I was surprised to learn (from this article) that there are local models that can do this (not sure if there are any that run on hardware I actually have though, unlike Tesseract which works fine on the scanning hardware I set up for it ~5 years ago.) For privacy reasons, cloud-based OCR is a non-starter...

Quite, I threw a so-so photo of an old, long receipt at Qwen 3.5 0.8MB (runs in <2GB) and it nailed spitting 20+ items out in under a second. AI is good at many things, but picking modern dependencies not so much.