How does your accuracy compare with VLMs like ColFlor and ColPali?
replies(1):
Here’s a video showing it in action: https://www.loom.com/share/ae43b6abfcc24e5b82c87104339f2625?..., and a demo playground (no signup required!) to test it out: https://app.midship.ai/demo
We started 5 months ago initially trying to make an AI natural language workflow builder that would be a simpler alternative to Zapier or Make.com. However, most of our users seemed to be much more interested in the basic (and not very good) document extraction feature we had. Seeing how people were spending hours a day manually extracting data from pdfs inspired us to build what has become Midship!
The problem is that despite all our progress in software, huge amounts of business data still lives in PDFs and images. Sure, you can OCR them, but getting clean, structured data out is still painful. Most existing tools just give you a blob of markdown - leaving you to figure out which parts matter and how they relate.
We've found that combining OCR with language models lets us do something more useful: extract specific fields and tables that users actually care about. The LLMs help correct OCR mistakes and understand context (like knowing that "Inv#" and "Invoice Number" mean the same thing).
We have two main kinds of users today, non-technical users that extract data via our web app and developers who use our extraction api. We were initially focused on the first one as they seemed like an underserved part of the market, but we’ve received a lot of interest from developers who face the same issues.
For pricing, we currently charge a monthly Saas fee per seat for the web app and a volume based pricing for the API.
We’re really excited to share what we’ve built so far and look forward to any feedback from the community!
Firstly as a function of the independent components in our pipeline. For example, we rely on commercial models for document layout and character recognition. We evaluate each of these and select the highest accuracy, then fine-tune where required.
Secondly we evaluate accuracy per customer. This is because however good the individual compenents are, if the model "misinterprets" a single column, every row of data will be wrong in some way. This is more difficult to put a top level number on and something we're still working on scaling on a per-customer basis, but much easier to do when the customer has historic extractions they have done by hand.