I really like your local-first idea. Just one question: what do you think is still the biggest weakness of local OCR models compared with cloud APIs? Is it layout parsing, reading order, or messy scans?
The traditional OCR model can't understand context, and that's what cloud LLM did.
BTW, some local models are already quite good like MinerU. But it's too heavy to run on consumer-grade PC. So it's just hard to strike a balance.