Skip to content

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.

Solution blueprint

The problem

What happens without a platform

  1. 01Each department buys its own AI tool, so company data flows to several outside services under several sets of terms.
  2. 02Nobody can say which model saw which document, or who asked what.
  3. 03Per-token bills grow with use, are hard to forecast or charge back, and tie every application to one vendor’s API.
  4. 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.

  1. Layer 1

    People and applications

    • Staff chat workspace
    • Assistants and agents built on it
    • Developer API
    • Admin, audit and cost console
  2. Layer 2

    Gateway and policy

    • SSO and role-based access
    • Model routing by data class
    • Guardrails and PII redaction
    • Audit log and usage metering
  3. Layer 3

    Models and retrieval

    • Open-weight and licensed models
    • vLLM-class inference serving
    • Embeddings and vector store
    • Arabic and English evaluation suite
  4. 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

  1. Request sent
  2. Identity and data class checked
  3. Policy selects the model
  4. Model answers, guardrails check
  5. 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

  1. 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.

  2. 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.

  3. 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.

  4. 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.

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.

Newsletter

Practical AI, in your inbox.

Real use cases, Saudi regulations and what works in Arabic — once or twice a month. No spam; unsubscribe any time.

We’ll e-mail you to confirm. Unsubscribe any time — see our privacy policy.