I build private Arabic AI for Saudi companies.
Assistants that answer your staff’s questions from your own documents (policies, contracts, procedures) and run inside your own cloud account in a Saudi region. Your data never leaves it.
Muhammad Uzair
AI engineer. I train models on private data and run them where the data lives.
- Question
- وفق نظام العمل، ما نسب مكافأة نهاية الخدمة المستحقة للعامل عند الاستقالة بحسب مدة خدمته؟ وفي أي الحالات تستحق المكافأة كاملة استثناءً من ذلك؟
- Answer
- يستحق العامل ثلث المكافأة عند انتهاء علاقة العمل بسبب استقالته إذا كانت خدمته لا تقل مدتها عن سنتين متتاليتين ولا تزيد على خمس سنوات، ويستحق ثلثيها إذا زادت مدة خدمته على خمس سنوات متتالية ولم تبلغ عشر سنوات، ويستحق المكافأة كاملة إذا بلغت مدة خدمته عشر سنوات فأكثر، وفقاً للمادة (85) من نظام العمل.
- Cites
- المادة (85) من نظام العمل
- Graded
- Passed: 5 of 5 required points, right article
One of 286 graded cases from Nizam, a Saudi labour-law assistant I built. Copied word for word.
Why most companies still don’t use AI
The data can’t leave
Company documents, customer records and employee data fall under PDPL and your own policies. Sending them to a public AI service is often not allowed.
General AI is unreliable on your specifics
On 50 Saudi labour-law questions answered from memory, the best general model I tested (Gemini 3.1 Pro) was fully right 78% of the time; GPT-5, 48%. And neither has ever seen your internal policies.
Nobody knows where to start
Which model, which cloud, what it costs each month, and who keeps it running. Most teams stall here.
What it looks like in your company
The assistant reads the documents your staff already search through, and answers with the source.
Reads
Carrier contracts, service-level agreements, customs procedures, driver and safety policies.
Staff ask
ما الغرامة إذا تأخر الناقل في التسليم وفق عقدنا المعتاد؟
What is the penalty if a carrier delivers late under our standard contract?
ما المستندات التي تطلبها الجمارك لشحنة إعادة تصدير؟
Which documents does customs need for a re-export shipment?
Illustrations of what a pilot would test, not past client work.
What I build for you
- An assistant on your own documents
- Staff ask in Arabic or English. It answers from your documents, shows the source, and says “not covered” instead of guessing.
- Running in your own cloud account
- Deployed in a Saudi region of the cloud you already use. You own the servers, the keys and the logs.
- Tested on your questions first
- Before you commit, I measure it on your real questions, the same way I tested Nizam, and show you what passes and what fails.
Case study: Nizam, built to prove the method
Nizam is a demonstration project, not a product I sell. I built it to prove the method on a hard Arabic domain, and the same steps work on your own documents. Ask it about Saudi labour law in Arabic: it answers from the official text, names the article, or says plainly that the texts don’t cover it. I trained it on ALLaM-7B, the Saudi Arabic model built by SDAIA and now published by HUMAIN, using about 5,400 reviewed examples written from 846 official articles.
| Measure | ALLaM-7B, untrained | Nizam v1 |
|---|---|---|
| Answers fully correct | 16% | 79% |
| Cites the right article | 25% | 98% |
| Says so when the texts don’t hold the answer | 2 of 30 | 27 of 30 |
| Refuses although the answer is there | 2 of 256 | 0 of 256 |
Graded by fixed rules, not by another AI. An answer passes only if it states every required point, cites the right article and declines when it should. No test question was used in training.
What it still gets wrong
- About 1 answer in 5 misses a condition or point.
- It only knows the texts it is given; it does not know later amendments.
- It is built for use where a qualified person checks each answer, not for legal advice.
What this means for cost
Better data moved the model; bigger hardware did not. The final version was trained on a single GPU. A useful private model does not need a data centre.
How a pilot works
Three steps. You decide after each one whether to continue.
- 1
20-minute call
You describe the use case. I tell you honestly whether a private model can do it, even if the answer is no.
You can stop here
- 2
Accuracy test on your questions
Fixed scope and fixed price, agreed before we start. You get a report of what it gets right and wrong.
You can stop here
- 3
Live in your cloud
If the numbers are good enough, it goes live in your account. You keep the model, the data and the setup.
Who holds what
You keep
- The servers, in your own cloud account in a Saudi region
- Your documents and every conversation
- The encryption keys and the logs
- The trained model: open models only, no lock-in
I get
- Only the access you approve
- Which you can revoke at any time
- With every action logged in your account
Designed with Saudi PDPL in mind. I hold no certifications yet (ISO 27001, SOC 2), so your security team can review the setup directly.
Who you would be working with
I fine-tune and deploy Arabic language models for narrow, high-stakes tasks, where a general chatbot is not good enough.
Before AI, I spent more than five years building production web systems, so I build the whole thing: the model, the tests, and the software around it.
- Experience
- Snyder Technologies, Full Stack DeveloperMarch 2023 – present, Karachi
- WizTech Pakistan, Full-stack DeveloperJuly 2021 – November 2022, Karachi
- Education
- University of the People, Bachelor's in Computer ScienceDecember 2024
- Tools
- Models: Fine-tuning (SFT, DPO, LoRA), Evaluation, Retrieval (RAG), vLLM
- Software: TypeScript, Python, FastAPI, Node
- Cloud: AWS, Cloudflare, Docker, Modal
I’m looking for 3 Saudi companies to pilot with in early 2027.
If your team wants to use AI but can’t send data outside, let’s talk for 20 minutes. You will leave with an honest view of what is possible, whether or not we work together.