Built to be checked.
Sovereignty is more than where the servers are. It is which AI models you use, who can see what, how personal data is handled, and a record of everything the platform did. This page is for the people who have to sign off on that.
- Where it runsDedicated to you: European cloud, your own cloud or on-premise.
- Your modelsChosen per agent, including your own.
- Who sees whatSearch enforces each person's access.
- Personal dataMasked before indexing. If it can't be masked safely, it is held back.
- EU AI ActProhibited uses are blocked when an agent is created and when it runs.
- AuditEvery action on record. Security events can't be switched off.
Where it runs
Dedicated to you, wherever you choose.
The whole platform is installed for you: the application, search, identity, and the models that read your documents. External services are used only where you choose them, such as an AI model provider.
European sovereign cloud
A dedicated deployment on European infrastructure, under European jurisdiction.
- Application
- Search
- Identity
- Indexing models
Your own cloud or data centre
Installed in your Kubernetes. With local models and self-hosted search, no external AI service is needed at all, so it can run air-gapped.
- Application
- Search
- Identity
- Indexing models
Hybrid, also supported
The platform in the cloud, the document reader inside your network next to file shares, SAP and databases. Raw documents stay home.
File shares · SAP · databases
Your models
Choose the AI, agent by agent.
huoku is not tied to one model vendor. Use a local model for sensitive work and a larger one where it's allowed, or keep everything in-house.
European-hosted and open-weight models through any OpenAI-compatible endpoint.
Your own models, served with vLLM or Ollama.
Or the major providers: OpenAI, Azure OpenAI, Anthropic, Google.
The models that read and index your documents always run inside the platform.
Who sees what
Permissions follow the data, from document to answer.
Access rules are applied inside the search engine, so an agent can never retrieve a document the person asking could not open.
Confluence, Jira, SharePoint, Slack and GitHub access rules are mirrored. Other sources get the access you assign.
Spaces keep departments separate. What belongs to another Space doesn't show up at all.
Automation runs as named service accounts, never as a person, and never with admin or builder rights.
Single sign-on with your identity provider, Microsoft group sync or LDAP.
Personal data
Masked before it is stored.
Data Guard masks personal data such as identity numbers and IBANs before anything is indexed. The original value never reaches the search index or the AI model.
- Tested live on the masking engine before it goes into effect.
- Every instance runs the same rules.
Fails closed: a record that can't be masked safely is held back, not let through.
Test patterns live before they go into effect. Changes reach every instance at once.
Finnish personal identity code included out of the box.
Also applied to the conversation summaries behind Memory.
EU AI Act
Prohibited uses blocked in the runtime.
Every agent declares what it is for. Uses prohibited under Article 5 of the EU AI Act are refused when the agent is created and again when it runs.
- Written as EU AI Act technical documentation.
- Every change creates a new snapshot, sealed with a hash.
Social scoring, workplace emotion recognition, manipulative techniques and the other Article 5 practices are blocked.
High-risk use cases are recorded.
Every agent gets a versioned, tamper-evident system card as technical documentation.
Human oversight through approvals on any action you choose.
Audit
A complete record, by design.
Who did what, when, with which agent. Security events are always recorded; conversation content only if you decide so.
- Prompt and answer text is not recorded by default.
- Filter by event, service or user, and export.
Logins, access changes, configuration, agent runs, tool calls, searches, email and more.
Authentication, access, configuration and admin events cannot be switched off.
Recording prompt and answer text is off by default and your decision.
Search, filter and export. Retention from six months to ten years.
For your technical team
The details your engineers will ask about.
Kubernetes and Helm. Infrastructure as code with a Terraform provider.
Single sign-on at the gateway. Encrypted, mutually authenticated traffic between services.
Rate limits per user and per agent. Your own certificate authority supported.
Security documentation and a technical walkthrough on request. No AGPL components in the document pipeline.
Bring your security team.
We'll walk them through the architecture, the controls and the audit trail on a live system.
Or write to sales@huoku.ai