One thing we noticed while building OCR is that accuracy drops heavily with low-resolution scans and rotated documents. GPU acceleration helped speed, but preprocessing ended up being the bigger challenge. Curious how others handle that.