Ai OCR

[autumn wyvern]

Like what do you actually want to build

Is it something like it reads the invoice and generates a report??

[barren wadi]

If you are looking for building a long term system, then you can create a basic web app where you'll upload your 100s or 1000s of invoices in batches/campaign and it will perform OCR to extract amount in your case and then you calculate total amount for that current batch using normal algorithm.

Apologies If you were not looking for this much technical response. 😅

[warped hull]

Nono, thanks, that's totally fine

[serene quail]

@warped hull Hi Daniel, I have the expertise of using Vision models and classic OCR system. We can have a talk.

[warped hull]

Hii

[serene quail]

@warped hull I have sent you a dm

[earnest tendon]

not sure if this is still going on but i built a fully functioning AI OCR platform both for generalized docuemnts to markdown and for invoice extraction. used qwen 2.5 vl instruction. About 5 seconds per page/invoice on a 12 gb 3060 using the 4b model

i tried everything... only thing that worked with scanned pdf's (images) was the vision models. It's not perfect but I created a json schema with validation (ie. line item data x quantity = line total?) duplicate invoices, etc. even wrote a POC front end that can be modified. currently just grabbing pdfs from a folder and running a timed or manual task.

current

process invoices for import takes the compiled json for each invoice, consolidates page page information then consolidates and exports to the ETL/CSV import format depending on the company. different companies use different vendor names, lists, codes, GL's and have different formats for entry.

have 2 a6000 ada cards coming that should let me process an invoice every 5 seconds or so. More than enough for our local deployment.

for context I'm integrating with sage 300 and Yardi Voyager 7/8

[winged burrow]

Hi @serene quail, I’m trying to build for a similar use case. Would love to connect.