# Sovereign AI

> Sovereign AI from Boltcliq: custom and fine-tuned models trained on your own data and deployed on premises, in your private cloud, on Boltcliq cloud or with a local provider, so you own the model, the data and the control.

Canonical URL: https://www.boltcliq.com/solutions/sovereign-ai
Part of: https://www.boltcliq.com/llms.txt

Own your AI. Custom and fine-tuned models, trained on your data and deployed where you decide: on your premises, in your private cloud, on ours, or with a local provider.
- Renting intelligence is not owning it. Your data, your model and your infrastructure should answer to you, in your jurisdiction, under your keys.
- Your model: Weights and adapters that belong to you
- Your data: Training and retrieval on data that stays under your control
- Your choice: On premises, private cloud, Boltcliq cloud or a local provider
- Your keys: Encryption and access decided by you, enforced by Odin
### What we do
- Most organizations use AI the way they use electricity: through a meter they do not control. Prompts, documents and customer records travel to a provider's servers, the model behind the service changes without notice, and the bill moves with someone else's pricing. For a bank, a ministry, a hospital or a telecom, that is a risk to manage, not a foundation to build on. Sovereign AI is the alternative: models you own, trained on data you control, running where you decide.
- The work has two halves. The first is making a model that is genuinely yours. We start from strong open or licensed base models, then adapt them to your language, your documents, your products and your rules. Depending on the problem, that means retrieval over your knowledge, supervised fine-tuning, parameter-efficient adapters, distillation into smaller and cheaper models, or a combination. We evaluate every step on your own cases, so you can see where a custom model beats a general one and where it does not.
- The second half is deployment. A model is only sovereign if you control where it runs. We deploy fine-tuned models on your premises, including air-gapped sites, inside your own cloud account, on Boltcliq-managed private cloud, or with a local cloud or data centre provider you choose. The same model and the same controls move between these options, so a choice made today does not lock you in tomorrow.
### Four things you can own
- Sovereignty is not a single switch. It is four separate questions, and each can be answered differently. Who owns the data used to adapt the model? Who owns the resulting weights? Who controls the infrastructure it runs on? And whose law and whose keys govern it?
- We design the answer to each, write it down, and build it so it can be checked. Which obligations apply depends on your sector and countries, so we work alongside your counsel and do not give legal advice.
- Data you control end to end
- Model weights and adapters you can take with you
- Infrastructure in a location you choose
- Keys, access and audit under your policy
### Capabilities
**Fine-tuning on your own data**
- We adapt open and licensed models to your domain using supervised fine-tuning and parameter-efficient methods. Training data is prepared, cleaned, de-identified where needed and versioned. Training runs inside your boundary, so raw data does not leave it. You receive the adapters, the training recipe and the evaluation report.
- Data preparation and de-identification
- Fine-tuning and adapter training
- Reproducible training recipe and reports
- A model that speaks your terms makes fewer costly mistakes.
**Retrieval and knowledge grounding**
- Not every problem needs training. Often the right answer is a general model grounded in your documents, records and ontology at question time. We build retrieval over your knowledge with access controls, citations and freshness rules, and combine it with fine-tuning when the evidence says it helps.
- Retrieval with citations and permissions
- Ontology-aware context through Zeus
- Fine-tuning only where it earns its cost
- Fresh, traceable answers without retraining every week.
**Small, efficient and local-language models**
- Bigger is not always better. We distil and quantise models to run on modest hardware, including a single server or an edge site, and adapt them for the languages your customers actually speak. This cuts cost, reduces latency and makes on-premises deployment practical.
- Distillation and quantisation
- Local-language adaptation and evaluation
- Right-sized models for each task
- The cheapest model that does the job is the sovereign one.
**Evaluation you can trust**
- We build an evaluation set from your real cases and use it to compare base models, tuned models and prompts. We test for accuracy, refusals, bias, leakage of private data and behaviour in local languages. Results are repeatable, so you can re-run them whenever the model or your data changes.
- Task-specific evaluation sets
- Safety, bias and leakage tests
- Regression checks on every release
- You cannot own what you cannot measure.
**Deployment where you decide**
- We deploy to four kinds of place. On your premises, including air-gapped sites. In your own private cloud account. On Boltcliq-managed private cloud, single tenant. Or with a local cloud or data centre provider you choose. Argus provides the portable runtime, so the model, its controls and its monitoring move between them.
- On premises and air-gapped
- Your private cloud account
- Boltcliq-managed private cloud
- Local provider or data centre
- Location is a policy decision, not a vendor default.
**Governance, keys and audit**
- Every model call is authenticated, authorised and recorded. Odin issues identities and credentials, Monolith sets what each agent and model may see and do, and Cyclopes keeps the evidence. Encryption keys are held by you. Prompts, outputs and logs follow the same placement rules as the data they came from.
- Customer-held keys
- Per-agent permissions and approvals
- Audit trail for every model call
- Sovereignty without control is only a location.
### Where it applies
- Banks and fintechs
- A bank wants an assistant for credit analysts and customer service, but customer records cannot be sent to a foreign-hosted model, and regulators ask where the model runs.
- We fine-tune a model on de-identified internal documents and deploy it inside the bank, on premises or in its private cloud, with keys held by the bank and every call logged.
- Risk and compliance decide which tasks the model may handle and what needs human approval. Counsel confirms the regulatory position.
- Government and public sector
- A ministry wants an assistant for citizens and civil servants in several local languages, but sensitive data must stay within national infrastructure.
- We adapt a model to local languages and official documents and deploy it in-country with a local provider, under the ministry's identity and audit controls.
- Officials decide the scope, the languages and the escalation rules. Oversight bodies review the evaluation results.
- Healthcare
- A hospital group wants clinical documentation support, but patient data is highly sensitive and connectivity between sites is unreliable.
- We deploy a compact model on premises at each site, grounded in the hospital's own protocols, with de-identified feedback used for improvement under clinical governance.
- Clinicians review every output before it enters a record. Governance boards approve changes.
- Telecommunications
- An operator wants customer-care and network-operations assistants trained on its own tickets and manuals, and wants to offer similar services to enterprise customers.
- We fine-tune and serve models on the operator's infrastructure, with metering by tenant so enterprise customers can be billed by use.
- The operator decides products, pricing and which customer data may be used.
### Built on
**Argus**
- Provides the portable runtime and private data platform where the model runs and its context is stored.
- Boltcliq product
- The model can move between sites without rebuilding.
**Monolith**
- Governs what the model may see, remember and do, and routes calls through one controlled gateway.
- Boltcliq product
- A sovereign model is also a governed one.
**Odin**
- Issues identities and credentials, enforces access and holds the policy decisions for people and models.
- Boltcliq product
- Keys and access stay under your policy.
**Cyclopes**
- Records inputs, outputs, cost and quality so every model call can be explained and re-tested.
- Boltcliq product
- You can show a regulator what the model did.
### What we do and do not do
- Fine-tune, distil and evaluate models on your data
- Deploy on your premises, your private cloud, Boltcliq cloud or a local provider
- Hand over weights, adapters, recipes and evaluation sets
- Design key control, access and audit with your counsel
- Train your operators to run and improve the model
- Train other customers' models on your data
- Claim a tuned model always beats a general one without testing
- Guarantee GPU supply or a specific chip
- Give legal advice on data protection law
- Name providers or partners without an agreement in place
### Questions we hear
- What does sovereign AI mean?
- It means your organization controls the data, the model, the infrastructure and the keys behind your AI, in the jurisdiction you choose, instead of depending on a service whose terms and location someone else sets.
- Do we need to fine-tune a model?
- Not always. Many tasks work well with a general model grounded in your own documents. We test retrieval, fine-tuning and distillation against your cases and recommend the lightest approach that meets the bar.
- Who owns the fine-tuned model?
- You do, subject to the licence of the base model, which we check with you before we start. You receive the adapters or weights, the training recipe and the evaluation set. We do not use your data to train anyone else's model.
- Where can the model run?
- On your premises, including air-gapped sites. In your own private cloud account. On Boltcliq-managed private cloud. Or with a local cloud or data centre provider you choose. The same model and controls move between them.
- Which models do you start from?
- Open-weight and licensed models for language, vision and speech, chosen for your task, languages and licence needs. We recommend after testing on your own cases.