Qelv

A prototype we built for an insurance operation in Thailand. 2026.

46 fields out of a Thai insurance form, on video

An insurance operation in Thailand keys its forms into SAP by hand from scans in two languages, with stamps and handwriting across them. Gemini 2.5 Pro reads the form and returns the fields as JSON, and the recorded walkthrough shows it happening on the real form.

Insurance. Automate and Data.

Thai document extraction in figures

  • 46fields out of one scanned insurance form, as JSON
  • 2languages on the form, with stamps and handwriting across it

A Thai and English insurance form, read into named fields

Screens from this system are being added.

The situation

The operation's forms arrive as scans in two languages, with stamps and handwriting across them, and the fields are keyed into SAP by hand. Most extraction demos run on a clean English invoice. This one was built to behave on the real forms, and to show the team what the SAP entry stage would take once the fields were out.

What we built

  • A Gemini 2.5 Pro extraction pass over a sample Thai and English insurance form
  • A schema of named fields, returned as JSON rather than a wall of text, so the output can be checked against the form
  • A recorded walkthrough of the extraction, with the form and the response side by side
  • The next stage designed: validation rules on the extracted fields, then SAP GUI scripting through COM to key them in, mapped for the team to decide on

What changed

The operation saw its own form read into named fields on video rather than a clean English invoice, and had the SAP entry stage mapped well enough to decide on.

Read next

Where to next

Have a problem in this shape?

Say so in a few sentences. The reply comes from the person who built this one.

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.