Qelv

Back to Labs

Prototype

Multilingual document extraction

Extraction demos are always run on a clean, flat, English invoice. The documents that actually arrive are a photograph of a form in two languages with a stamp across the middle of it. This was built to behave on those.

What it shows you
How document extraction behaves on the hard inputs, not the tidy demo ones
On this page
No captures

This one is driven on a call.

Delivered to a client
Never

Built to settle a scope question. Nothing here is counted as work.

What is in it.

  • A Gemini extraction pass over scans, forms and mixed layouts, returning named fields rather than a wall of text
  • A schema per document type, so the output is checkable against something instead of being trusted
  • A confidence path: fields the model is unsure of are handed back for a person to confirm, not quietly filled in

Why there is no link.

It runs against documents supplied in presales conversations, which belong to the people who supplied them. There is no sample set we own that is representative of the hard inputs, and a demo on easy inputs would be the exact dishonesty this prototype exists to avoid.

Every figure inside these is sample data. We walk you through them live on a call; each card below has a booking link.

Elsewhere on the site

Related

Where to next

Want one of these built around your own workflow?

A working version in two weeks. Not a specification document, not a deck: something your team can click and argue with.

Send us the bottleneckAll prototypes

We reply within two business days. If it is a fit, you get a one-page note on what to do first, before any sales call.