I have been building document intelligence in enterprise for a long time now, and even with sota OCR, SLMs, LLMs, all the myriad of services now offering this - most of them miss something.

So over the years me (and my team) have built up a lot of experience dealing with these for sensitive industries like healthcare/insurace/lending etc.

We are bringing these ideas out as a router to other parsers, but with some key takes that allow you to manage failures. "Failure is inevitable so route for it."

It is called Openreading - https://openreading.ai/ and open core here: https://github.com/openreading-ai/openreading-core

It is still WIP and will be more polished in a month after some more rigorous testing and benchmarks.

If this is a problem you deal with, would love to chat.

Note: the intent is to launch a managed version that provides more durable compute/retries, parallel execution, intent understanding etc but this is for later.