Sovereign LLMs, SLMs and context platforms
Language models that answer from your own data and never send it outside your control
Your data cannot go to a public AI service
Banks, ministries and other regulated institutions cannot send customer or citizen records to public AI services abroad. Residency laws forbid it, and so do most risk committees.
Yet the same institutions need assistants, search and drafting that understand their own documents, often in Arabic as well as English. A model that cannot see the data is of little use, and one that ignores permissions is a risk.
What you get
Data stays in the country
Models that run inside your own data centre, private cloud or a regional sovereign cloud, so data stays in the country.
No training on your data
No third party trains on your prompts, documents or answers.
Permissions respected
Answers built only from what the person asking is allowed to see.
Arabic and regional dialects
Models tuned for Arabic and regional dialects where your users need them.
What we deliver
Air-gapped, on-premises and sovereign-cloud hosting
Models served in an air-gapped environment, on your own GPU servers or in a private network on a regional cloud, sized to your workload.
Frontier and open-weight models
The right model for each task, from OpenAI, Claude and Gemini to open-weight Llama, Qwen, DeepSeek and Mistral, behind one interface so it can be replaced.
Arabic-first and regional fine-tuning
Small language models such as Jais, Falcon and ALLaM, and European and Asian models, fine-tuned on your vocabulary, dialects and documents.
Context platforms and permission-aware GraphRAG
A semantic layer and knowledge graph over your content, with retrieval that applies row-level and role-level access rules before the model sees any text.
How it works
Model hosting runs on on-premises GPU clusters or air-gapped private clouds, with Jais, Falcon, ALLaM, Mistral, Qwen, DeepSeek or Llama, and access rules enforced before inference.
Choose the job and the boundary
We agree who will use the model, the tasks it handles, the data it may read and where it is allowed to run.
Prepare the data
We bring the sources into a governed data layer and map the permissions that must carry through to every answer.
Select and tune the model
We compare models on your own questions for quality, language, speed and cost, and fine-tune a small model where it pays.
Deploy inside your boundary
We install the model and retrieval in your environment, with access control, logging and security review.
Measure and operate
We track quality, cost and usage, review flagged answers with your team and keep the evaluation questions growing.
Technologies and standards
The data flows from your sources through an Apache Iceberg or Databricks layer and permission-aware GraphRAG into a model in an air-gapped private network. Data never leaves the enterprise.
- AI and machine learning
- OpenAI GPT
- Anthropic Claude
- Google Gemini
- Meta Llama
- Mistral
- Hugging Face
- Amazon Bedrock
- Azure OpenAI
- Vertex AI
- Data and analytics
- Databricks
- Methods and practices
- Permission-aware GraphRAG
- Semantic layers
- Air-gapped private networks
- Small language models (SLMs)
How to start
Most clients start with the sovereign agentic proof of concept: a working, permission-aware GraphRAG or agent workflow connected to your internal tools in a private network.
What happens next
Agreeing a time
Our team emails you to agree a time for the session.
Pre-consulting session
A free 30-minute online pre-consulting session: we go through your goals, systems and constraints.
Readiness review
In a free 48-hour readiness review, a Principal Architect checks your project against the 6-point Definition of Ready and recommends the track that fits.
Choosing the track
Scope-ready projects receive a binding proposal within 3–5 business days. Concept-stage projects start in the Blueprint Studio. A live system that is failing, undocumented or needs integration starts with a diagnostic of 1–2 weeks. The Blueprint Studio and the diagnostic each have a fixed fee, quoted in the proposal.
AI-native product teams
Managed, cross-functional pods of senior engineers, equipped with agentic coding and QA harnesses.
Fixed-outcome projects and round-the-clock managed services
Fixed-scope, milestone-driven delivery with named delivery leads, and round-the-clock application managed services under SLA governance.
The quickest way in is the Sovereign agentic POC (MCP, A2A, GraphRAG) package, one of our rapid engagement packages.
Questions and answers
Can the model run without any data leaving our country?
Yes. We host the model in an air-gapped environment, on your own servers or in a private network on a regional sovereign cloud. Prompts, documents and answers stay inside that boundary, and no third party trains on them.
Which language models do you work with?
We choose per task. Options include frontier models from OpenAI, Claude and Gemini, open-weight Llama, Qwen, DeepSeek and Mistral, and regional models such as Jais, Falcon and ALLaM. Where data must stay in the country, we use models that can run inside your environment.
Does it work in Arabic and regional dialects?
Yes. We fine-tune Arabic-first small language models such as Jais, Falcon and ALLaM on your vocabulary and documents. We then test them on the questions your users ask, in the dialects they use.
How do you stop the model showing people what they should not see?
Retrieval applies your row-level and role-level access rules before any text reaches the model. An answer is built only from content the person asking may read, using your existing identity and access rules.
What is permission-aware GraphRAG?
Retrieval-augmented generation finds the passages that answer a question and gives them to the model. GraphRAG adds a knowledge graph of how your entities relate, so answers can follow those links. Permission-aware means every passage is filtered by the asker’s rights first.
Do we need a large model, or will a small one do?
Often a small one will do. Classification, extraction and answers from your own documents usually work well on a tuned small language model, which runs on fewer GPUs and costs less. We keep large models for the tasks that need deeper reasoning.
Who owns the fine-tuned model?
You do. The code, prompts, evaluation sets and any fine-tuned model weights we produce for you are yours once each milestone is paid. Commercial and open-weight base models remain under their own licence terms, which we list for you.
Start with a conversation, not a contract.
Tell us what you need to build, connect or secure. An architect goes through it with you, then recommends where to start.