Platform

A complete AI stack, from GPUs to applications, inside the Kingdom.

Seven layers, each one running in-Kingdom and each one under your control: what runs, which version, who can call it and what was logged.

منصة ذكاء اصطناعي متكاملة داخل المملكة

Architecture

Seven layers, one boundary

Requests enter at the top and never leave the Kingdom on their way down. Layer 4 applies your policies to every request and every response.
  1. 7

    Applications

    Assistants · document Q&A · drafting · contact-centre copilots · code assistants

    Ready-made workspaces for staff, or your own applications on the API.

  2. 6

    Agents & workflows

    Tool calling · multi-step workflows · human approval steps

    Agents act only through tools you register, with an approval step before anything leaves the platform.

  3. 5

    Retrieval

    Connectors · Arabic and English search · vector store · citations

    Answers grounded in your documents, with the passage cited. Permissions from the source system are respected, so people only retrieve what they may already read.

  4. 4

    Guardrails & evaluation

    PII redaction · policy filters · prompt-injection defence · evaluation sets

    Policies applied on the way in and out, and every model release tested on your own evaluation set before it is promoted.

  5. 3

    Inference gateway

    OpenAI-compatible API · routing · quotas · usage metering

    One endpoint for every model. Existing code changes a base URL and a key, not its logic.

  6. 2

    Model hub

    Arabic-first and open-weights models · fine-tuning · version pinning

    Curated models served in-Kingdom, plus your own fine-tuned or uploaded weights.

  7. 1

    Sovereign infrastructure

    In-Kingdom GPU compute · storage · HSM-backed keys · isolated networks

    SoverAIn Node, Server or Rack: NVIDIA GPU hardware in your facility or an in-Kingdom data centre. The controls above are the same on all three.

Capabilities

What you can build on it, and what is ready when

Labels are commitments. Design-partner capabilities are being built with our first customers now.

Chat and completion

Arabic and English assistants for staff, with conversation history kept in-Kingdom.

Design partners wanted

Document intelligence

Read scanned and digital documents in Arabic and English; extract fields to a schema; summarise and compare.

Design partners wanted

Grounded search (RAG)

Question answering over your documents with citations and source permissions respected.

Design partners wanted

Embeddings

Arabic-aware embeddings for search, de-duplication and classification.

Design partners wanted

Fine-tuning

Adapt a model to your terminology and tone on your data; the resulting weights stay in your tenancy.

Early access

Agents

Multi-step workflows that call registered tools, with human approval before any external action.

Early access

Speech

Arabic speech-to-text across Gulf dialects and text-to-speech for contact centres.

On the roadmap

Air-gapped deployment

SoverAIn Node or Server inside your own facility, with no external connection at all.

Design partners wanted

Model hub

Open models you can inspect, pin and adapt

No single model is best at everything. The hub serves a curated set, and the right one for each task is chosen on your own evaluation set, not on a public leaderboard.
FamilyExamplesTypical use
Arabic-firstOpen models trained for Arabic, such as ALLaMArabic drafting, summarisation, correspondence, citizen and customer services
General open-weightsFamilies such as Llama, Qwen and Mistral, at several sizesReasoning, extraction, coding, bilingual work
EmbeddingsMultilingual embedding models with Arabic supportSearch, retrieval, clustering
Your ownYour fine-tuned or uploaded weightsDomain language, house style, specialist tasks

Version pinning

A model version never changes under you. Upgrades are tested on your evaluation set and promoted by you.

Fine-tuning stays home

Adapters and weights trained on your data live in your tenancy and are deleted on request.

External models, by exception

An offshore model can be enabled for a workspace only by an administrator, in writing, with a banner on every response.

Model names are examples of families we evaluate. Availability of a specific model depends on its licence and on your evaluation results.

API

Keep your code. Change the base URL.

The inference gateway speaks the OpenAI-compatible API, so existing SDKs, frameworks and applications move across with a configuration change.
curl https://api.soverain.cloud/v1/chat/completions \
  -H "Authorization: Bearer $SOVERAIN_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "arabic-instruct",
    "messages": [
      { "role": "user", "content": "لخّص هذا الخطاب في ثلاث نقاط" }
    ]
  }'

# Same request shape as the OpenAI API: point your SDK's base URL
# at SoverAIn and keep your code.
Illustrative request. Endpoint and model names are examples.

See it on your own documents.

A design-partner pilot runs one use case on your data, in the deployment model your classification requires.

Become a design partnerCompare deployment options