Case study · Dinarak

We built an AI that reads Dinarak's Arabic court notices field by field, and a person approves every wallet freeze.

Our AI reads and checks every field; anything under 95% confidence goes to a person, and only an approver can freeze a wallet.

Fintech, mobile wallet · Jordan

Your data stays where you choose: on your own servers, in your own cloud account, or with a contracted model provider. We confirm which in writing.

Court notice

Page 3 of 89

Scanned page

1234

Read field by field

1Name
99%
A. Example
2National ID
98%
9•••••••312
3Wallet
97%
WLT-••••-4821
4Case number
91%
2026/•••8
Under 95%: sent to a person
Freeze walletWaits for an approver
Example screen, example data

Results

5,600+
pages of court notices read
Counted in the system's job records
17,000+
debtor records extracted
Counted in the system's records
591
wallet freezes confirmed
Confirmed by the system's follow-up check

Watch it work

Court notice

Example
  1. A scanned court notice arrives · 89 pages, Arabic

  2. Reading page 3, field by fieldNameNational IDWalletCase number
  3. Checking the ID format, the name against the wallet, duplicates and the account type
  4. Looking up the walletWallet system
  5. The case number was read at 91%, under 95%, so it goes to a person

  6. Confirmed by a reviewer

  7. The freeze waits for an approver

  8. Approver's decision

    Only an approver can freeze the wallet

What the AI does
Where it stops for a person
What it logs

Where a person decides

  • Any field read with less than 95% confidence goes to a person.
  • Only an approver can freeze a wallet, and real freezes stay off until an admin switches them on.
  • A freeze is blocked for companies, salary accounts, name mismatches, and IDs edited after the wallet check.

The challenge

Dinarak is a licensed mobile-wallet provider in Jordan. Court notices asking it to freeze customers' wallets arrived as scanned Arabic documents, and the team read each one by hand.

A wrong ID number or name freezes the wrong person's wallet, so every notice had to be read and checked carefully before anyone acted.

What we built

  • Reading of Arabic court notices, page by page and field by field.
  • Checks on every ID number, name and duplicate, a second read of any field that looks wrong, and a confidence score for each field.
  • A lookup in Dinarak's wallet system, the freeze itself, and a re-check that the freeze held.
  • A log of every action that can't be edited afterwards.

What we run after launch

  • We host it and watch it.
  • A daily check that every frozen wallet is still frozen.
  • A sweep for any job that didn't finish, and support for the team.
Next caseAqarati X: An Arabic-first marketplace with a dashboard for every role, and Dalia X, an AI assistant that explains how the platform works.

Facing something similar?

Tell us where the work piles up. If it fits, we start with Assess & Design, where we agree what to build and where a person decides.

Talk to us

You decide. We answer for it.

All case studies