AI Solutions

We build AI-powered websites, apps, agents and automations — and we specialise in Claude. You buy delivery capacity, not a vague retainer: a fixed number of engineering hours plus a token quota you can spend on any AI work you ask us for.

You see both meters in a monthly statement. Nothing is billed twice, and nothing is spent without a work item behind it.

Item — Pilot — Build — Scale — Embedded

Extra block of 10 hours — 760 — 670 — 600 — 560

Extra block of 10M tokens — 250 — 220 — 180 — 150

Rush delivery (48 hours) — +35% — +35% — +25% — —

Hourly, without a package — USD 95 / hour

Private or on-premise model deployment, and any work requiring a dedicated GPU host, is quoted separately. Unused tokens do not roll over on one-off packages; unused hours can be extended once by 30 days on request.

“Tokens” means nothing until you can price a job in them. These are real figures from work we have delivered — treat them as ranges, not quotes.

A job of this size — Typical tokens — Fits inside

Reading a 100-page tender and answering questions on it — 0.1 – 0.3M — Pilot, many times over

Extracting fields from 1,000 supplier invoices — 2 – 3M — Pilot

A support assistant handling 1,000 customer conversations — 4 – 6M — Build

One agent run across a medium codebase: read, plan, edit, test — 0.5 – 1.5M — Build ≈ 20 runs

Translating and localising a 50-page bilingual website — 1 – 2M — Pilot

An internal knowledge assistant used daily by 40 staff, for a month — 15 – 25M — Scale

A WhatsApp sales assistant, 300 conversations a day, for a month — 35 – 50M — Embedded

Arabic text uses roughly twice the tokens of the same content in English. We account for that in every quote — it is a known cost of working bilingually, not a surprise we pass on to you later.

A token is roughly three-quarters of an English word — the unit the AI model reads and writes in, and therefore the unit it is billed in. One million tokens is about 750,000 English words, or around half that in Arabic.

We tell you at 75% and again at 90%, before either meter runs dry — never after. You then add a top-up block or we pause. Nothing is consumed beyond your package without your written approval.

You do — all three, on final payment, including the prompts and the evaluation set. Your data is never used to train anyone's model. If you leave us, you leave with a system you can hand to another team.

Yes, for open-weight models, and we host it either in our facility or in yours. Frontier models such as Claude run as a hosted API — if a regulator requires everything to stay inside your walls, we say so at the scoping call rather than after you have paid.

Yes. Interfaces, prompts, evaluation sets and documentation, in Modern Standard and Gulf Arabic. Everything we ship exists in both languages from the first day, including this page.

Usually because nobody defined what “working” meant. That is what the evaluation set is for: an agreed test with a score, run before launch and re-run every month. If the score drops, we know before your customers do.

Forty-five minutes, no charge, no slide deck. You describe how the work is done today; we tell you what AI would change, what it would cost in hours and tokens, and where it would not help at all.