Private AI platform
One governed AI platform, running where you decide.
A private AI platform is the foundation every assistant and agent in your organisation runs on. We serve open-weight models, and commercial models licensed for private deployment, on GPU servers in your data centre or on dedicated private-cloud capacity in the Kingdom — behind one gateway that decides which data may reach which model. Your teams build on it; you keep control of the models, the data and the running cost.
The problem
What happens without a platform
- 01Each department buys its own AI tool, so company data flows to several outside services under several sets of terms.
- 02Nobody can say which model saw which document, or who asked what.
- 03Per-token bills grow with use, are hard to forecast or charge back, and tie every application to one vendor’s API.
- 04Arabic quality is taken from vendor claims instead of being measured on your own documents and questions.
How it is built
How the platform is built
Four layers, each replaceable on its own. Applications never call a model directly: every request passes through the gateway, where identity, data classification and policy decide what happens next.
Layer 1
People and applications
- Staff chat workspace
- Assistants and agents built on it
- Developer API
- Admin, audit and cost console
Layer 2
Gateway and policy
- SSO and role-based access
- Model routing by data class
- Guardrails and PII redaction
- Audit log and usage metering
Layer 3
Models and retrieval
- Open-weight and licensed models
- vLLM-class inference serving
- Embeddings and vector store
- Arabic and English evaluation suite
Layer 4
Infrastructure
- GPU servers on-premises
- Or private cloud in the Kingdom
- Container platform and secrets
- Metrics, traces and alerts
Capabilities
Private AI platform
A model gateway with policy
Every request goes through one gateway that knows the user, the application and the data classification. Policy decides which model may receive it — for example, confidential data only to models hosted on your own servers.
Open-weight models, served properly
We serve open-weight models such as Llama, Qwen or Mistral, and Arabic-first models such as ALLaM or Jais where the licence allows, on vLLM-class inference servers sized to your load. External commercial APIs can sit behind the same gateway for the data classes you approve.
Quality measured, not assumed
An evaluation suite tests each candidate model against sets built from your own tasks, in Arabic and English. Any change to a model, prompt or setting is compared with the last results before it goes live.
Guardrails and redaction
Personal data such as national ID and iqama numbers, IBANs and mobile numbers can be detected and masked before a prompt leaves a trusted zone. Input and output filters apply your acceptable-use policy the same way in every application.
Identity, audit and metering
Single sign-on and role-based access decide who may use which model. Every request is logged with user, application, model and token counts, so usage can be reviewed and its cost charged back to departments.
Freedom to change models
Applications call one stable API, not a particular model. When a better or cheaper model appears, you evaluate it, change the route and leave the applications as they are.
How it works
How a request travels
- Request sent
- Identity and data class checked
- Policy selects the model
- Model answers, guardrails check
- Logged and metered
Connects to
Works with
- Microsoft Entra ID and Active Directory
- Okta and other SAML or OIDC identity providers
- Llama, Qwen and Mistral open-weight models
- ALLaM and Jais Arabic-first models
- Commercial model APIs, where policy allows
- vLLM and other inference servers
- Kubernetes and OpenShift
- PostgreSQL with pgvector, Qdrant or Milvus
- Splunk, Elastic and other SIEM platforms
- Prometheus and Grafana
Where it runs
On-premises
Models and data run on hardware in your own facilities. Nothing leaves your network unless you allow it.
Private cloud in the Kingdom
Dedicated, isolated capacity in a Saudi data-centre region of a provider you select — the elasticity of cloud without a shared public AI service.
Hybrid
Sensitive workloads stay on-premises; less sensitive ones use private cloud capacity. You set the policy for what goes where.
What we measure
Measures we agree with you
We agree these measures with you during discovery and report against them once the platform is running. They are targets we track together, not results we promise in advance.
01
Evaluation score by task
Accuracy on your Arabic and English test sets for each use case, re-run whenever a model changes.
02
Cost per request
Total running cost divided by requests served, set against the per-token price of an equivalent public service.
03
Latency and availability
Response time in the busiest hour and platform uptime, against levels set for each application.
04
Share of governed usage
How much of the organisation’s AI use runs through the platform, with its policies and logs, rather than through unmanaged tools.
Indicative delivery path
Indicative delivery path
- 01
2–4 weeks
Assess and size
We list use cases and data classes, test candidate models on samples of your Arabic and English content, and size GPU or private-cloud capacity with a running-cost estimate.
- 02
4–8 weeks
Build the foundation
We deploy serving, the gateway, identity integration, logging and the evaluation suite in your environment, then harden it with your security team against your policies and the NCA Essential Cybersecurity Controls you apply. Hardware lead times, if you are buying servers, come on top.
- 03
3–6 weeks
First workloads
The first one or two applications — often an internal chat workspace and a knowledge assistant — go live on the platform, with agreed measures.
- 04
Ongoing
Operate and extend
We monitor quality, cost and capacity, evaluate new models as they are released and bring further use cases onto the platform.
Durations are indicative and depend on scope, data and integrations.
Questions
Private AI platform
Is running our own models cheaper than paying per token?
It depends on how much you use and how steady that use is. GPU servers are a large upfront cost and need people to run them; per-token services cost nothing when idle but grow with every request. We model both against your expected load before you commit, and a hybrid route is often the sensible middle.
Are open-weight models good enough in Arabic?
Some are, for some tasks — which is why we measure rather than assume. We test candidates on your own documents and questions in Arabic and English, including dialect where it matters, and choose per task. Where no open model is good enough, policy can allow an approved commercial model for the data classes you choose.
Who operates the platform after launch?
Your team, with our training and runbooks; Aahlan, under a managed-operations agreement; or a mix of the two. Running GPUs, updating models and covering on-call support is real work, and we are clear about what it takes before you decide.
Does hosting in the Kingdom make us PDPL-compliant?
It helps with where personal data is stored and processed, but it is not the whole of compliance. The Personal Data Protection Law also covers purpose, consent, retention and data-subject rights across your organisation. We design the platform to support those obligations and work alongside your legal and compliance teams; we do not certify compliance.
Other solutions
All solutionsKnowledge assistant
Cited answers from your documents and systems, in Arabic or English — limited to what each user may open.
View solutionDocument intelligence
Invoices, contracts and applications read, checked against your records and posted — people review the rest.
View solutionCustomer service agent
Arabic and English service on web chat, WhatsApp and voice — approved answers, lookups, hand-over to people.
View solutionAgentic operations
AI agents that handle procurement, invoice, HR and IT requests end to end — within limits you set.
View solution
Have a process that should run itself?
Tell us about it. We will help you work out whether AI is the right answer, what it would take, and what it should return.

