AI solutions

Arabic AI inside the work you already do.

Our models read printed and handwritten Arabic, turn speech into text, and find the passage that answers a question. Each is one step in a process a person approves.

  • Models we chose, tuned and measured ourselves
  • A person approves every AI step
  • Served on our GPUs in Saudi Arabia

One pattern

Five workflows, and a person closes every one.

The AI step sits in the middle. It suggests, reads or listens; a person accepts the result.

  1. A contract is registered

    1. Uploaded
    2. AI step: Details suggested
    3. A person approves
  2. A newspaper page is digitized

    1. Boxed
    2. AI step: Read
    3. A person corrects
  3. An interview is archived

    1. Recorded
    2. AI step: Transcribed
    3. A person reviews
  4. A photo is described

    1. Added
    2. AI step: Weak description flagged
    3. A person improves it
  5. A request for a rare item

    1. Received
    2. AI step: Matched to holdings
    3. A person confirms

How we build a model

Choose, score, fine-tune, host.

  1. 01

    Choose

    We test the open and commercial options side by side before we pick one.

  2. 02

    Score

    A fixed exam with 95% intervals, so a new version is compared only on the same test.

  3. 03

    Fine-tune

    Handwriting gets its own model and print keeps its own, so one never degrades the other.

  4. 04

    Host

    Served on our own GPUs in Saudi Arabia, next to the systems that use them.

Kindi Labs

Try the models yourself.

Create an account in Kindi Labs, upload a page, and read the output line by line: text, boxes and JSON.

Questions we hear

Is this a chatbot?

No. AI is one step inside your existing workflow: it suggests, reads or listens, and a person approves the result.

Where do the models run?

On our own GPUs in Saudi Arabia, next to the systems that use them. Documents and recordings are handled by a local model, with no outside API involved.

How do you know a model is good enough?

We score every version on a fixed test with 95% intervals and publish the numbers, including where our lead isn't clear of the next system.

Can we try it before we talk?

Yes. The Kindi Labs playground lets you upload a page and see what the models read.

Have an archive, a form, or recordings to read?

Tell us what comes in and who checks it. We'll show where a model fits and how it would be measured.

Start a project