How Diana works. No hand-waving.
A plain-language explanation of the sovereign architecture — for technical buyers, IT directors, and compliance officers who need specifics, not marketing.
Choose your sovereignty level.
Dedicated EU server. No shared tenancy.
Diana runs on a dedicated private server in a European data centre — OVHcloud, Hetzner, or Scaleway. Your data stays on EU soil, on hardware that only you use. No public cloud. No data commingling.
Best for: Mid-size firms, agencies, boutique practices
Deployment: Live in 24 hours
Price: From EUR 5,000/month
Hardware in your IT room. Your data never leaves.
Diana ships as a physical appliance. Your team plugs it in. We configure and support it remotely through a secure management channel that has zero access to your data. Everything runs locally.
Best for: Law firms, investment banks, private equity funds
Deployment: Live within 48 hours of delivery
Price: From EUR 10,000/month + setup
Zero network connection. Absolute sovereignty.
For environments where even a local network connection is too much. Diana ships pre-loaded. No internet. No management tunnel. No external dependency of any kind. Data physically cannot leave.
Best for: Government, defence-adjacent, sovereign wealth funds, classified-adjacent legal work
Deployment: By arrangement
Price: Custom
The intelligence layer
What is actually running inside the box.
Diana's agents are powered by Small Language Models — a category of AI designed to run on compact, dedicated hardware rather than requiring cloud-scale data centres. This is the architectural decision that makes everything else possible. Without SLMs, sovereign on-premise deployment would not perform at the speed professionals need. With them, a box the size of a small desktop delivers enterprise-grade AI execution at 25-second workflow speeds.
Unlike large public models that are trained to handle any question from any person, Diana's SLMs are fine-tuned on specialist expertise. Each agent is trained by a domain expert through the Expert Marketplace — encoding their specific methodology, frameworks, and decision logic into a model that thinks like them. The result is not general intelligence. It is expert intelligence. Exactly the kind professional services requires.
Two pipes. Two purposes. They never touch.
Data traffic and management traffic are completely separate systems.
The data channel — stays inside forever.
Your employees' requests, documents, CRM data, and workflow outputs travel from their laptop to Diana's server through your internal network. This channel has no connection to the internet. Data physically cannot leave.
The management channel — carries zero data.
A separate management controller connects to Diana's team through an encrypted tunnel. We push software updates, monitor server health, and fix issues remotely. This channel has no access to your data — by architecture, not by policy.
What Diana can do remotely
- Monitor server health (CPU, memory, storage, uptime)
- Push software updates and new agent versions
- Diagnose and fix technical issues
- See aggregate workflow counts for billing (no content)
What Diana cannot do remotely
- Access any document, file, or workflow content
- Read inference results or agent output
- Access CRM, email, or legal documents
- View audit log content
How expert agents reach your infrastructure.
The agent travels to you. Your data never travels anywhere.
Subscribe
You browse the Expert Marketplace and subscribe to an agent — for example, an NDA Review Agent trained by a Magic Circle partner. Diana registers your subscription.
Secure delivery
The agent bundle — encrypted, signed, and tied to your hardware's unique certificate — downloads to your sovereign environment. Only your box can decrypt it.
Local execution
The agent runs on your hardware. It reads your documents, processes your data, executes your workflows. Everything happens inside your perimeter. Nothing goes back out.
Heartbeat only
A tiny encrypted signal — agent ran, count, timestamp — returns to Diana for billing. No content. No data. Just a count and a timestamp.
Technical FAQ
Can Diana access our data through the management tunnel?
No. The management tunnel connects to the management controller inside the box — a separate small device. It has no pathway to the inference engine, vector database, or any client data. Your IT team can verify this by inspecting the network configuration independently.
What type of AI model does Diana use?
Diana runs on Small Language Models — a category of AI designed to operate on compact, dedicated hardware rather than requiring cloud data centres. Unlike large public models that need to be accessed via an internet connection (which is why your data has to leave your environment when using tools like ChatGPT or Copilot), SLMs run entirely on the hardware inside your building. They are fine-tuned on specialist domain expertise rather than trained on general internet data, which makes them more accurate for professional tasks and categorically better suited to environments where data sovereignty is non-negotiable. The model processes your data locally. Nothing is transmitted. Nothing is shared.
What happens if Diana's company shuts down?
Your box keeps running. The software is on your hardware. Agents you have already downloaded continue to work. You are not dependent on Diana's servers for day-to-day operation — only for updates and new agent downloads. Your data was never on our servers. It stays exactly where it always was: inside your building.
How do software updates work without risking our data?
Updates are delivered through the management tunnel to the management controller — a separate small device inside the box. The update is a software package, not a data transfer. It moves from our update server to your box, installs in under 60 seconds, and restarts Diana silently. No client data moves in either direction during this process.
What is the difference between Sovereign Cloud and On-Premise?
Sovereign Cloud puts your dedicated server in a European data centre — faster to deploy, managed infrastructure. On-Premise puts it in your building — physical control, no external dependency. Both guarantee no shared tenancy and no data commingling. The decision comes down to whether you want the hardware in your building or in a certified European facility.
Can we audit the architecture ourselves?
Yes, and we encourage it. Your IT team can inspect the network configuration, firewall rules, management tunnel scope, and agent package verification process independently. We provide full technical documentation — architecture diagrams, network flow specifications, and security configuration guides — to every enterprise client before deployment.
How many users can Diana support?
A standard appliance (RTX 4070 Ti, 64GB RAM) supports 3–5 concurrent users comfortably. For larger teams, additional appliances are added to the same internal network and automatically load-balance. Each new box is identical in setup. Your team notices nothing except Diana getting faster. We size the deployment during your onboarding call.
Seen enough to talk?
Most deployments are live within 48 hours of your decision. No long procurement process. No complex setup. Just a box, your network, and sovereign AI running in your building.