NOVARA auf Xinity Runtime: KI-Kollege ohne Cloud-Anbindung

When Xinity emailed Novarion about co-writing this article, the first reply did not come from a human. It came from NOVARA, Novarion’s AI coworker. NOVARA reviewed the proposal, suggested a structure, and helped determine how this article should be framed. No external cloud API had to be involved in that process. That is the whole story in miniature: AI coworkers are no longer abstract pilot projects. They read email, support workflows, retrieve knowledge, and can operate close to an organization’s most sensitive information. The strategic question is therefore no longer whether AI coworkers are useful. It is where they run, how they integrate, and under whose control they operate.

For regulated organizations in the DACH region, that question is not philosophical. It is operational, legal, and architectural. If an AI system processes confidential business data, internal communications, customer records, project documents, or regulated information, sovereignty is not a marketing preference. It is a deployment requirement.

This article describes a reference architecture for exactly that requirement: NOVARA as the intelligence layer and Xinity Runtime as the local inference layer underneath it. Together, they show how an enterprise AI coworker can run on infrastructure that stays inside the customer’s own environment.

The problem

Regulated organizations want what everyone else already has. Banks want AI that reads credit files and supports internal knowledge work. Hospitals want AI that helps draft documentation and structure information flows. Law firms want AI that searches archives and assists professionals without weakening confidentiality. Industrial companies want AI that works across operations, service, engineering, and communication. The technology exists. In many cases, the real blocker is no longer capability.

The blocker is jurisdiction.

A large share of AI coworkers on the market depend on external cloud APIs for core inference. The moment prompts, documents, or contextual enterprise data leave the organization and are processed through an external endpoint, the compliance discussion changes fundamentally. For organizations bound by GDPR, banking secrecy, medical confidentiality, professional privilege, or internal sovereignty requirements, that architecture creates a level of exposure that contracts alone cannot fully neutralize.

This is why many organizations either postpone AI adoption or confine it to trivial, non-sensitive use cases. But that defeats the purpose of a real coworker. A true AI coworker is valuable precisely because it can work with the context that matters: enterprise systems, internal processes, and real operational data.

The answer, therefore, is not to make AI less capable. It is to make the stack sovereign. If the intelligence layer depends on an external API outside the organization’s control, sovereignty remains partial at best. A sovereign AI coworker needs sovereign infrastructure underneath. That combination is what this reference architecture is about.

The intelligence layer: NOVARA

NOVARA is Novarion’s Virtual Coworker for enterprise environments. It is designed not as an isolated chatbot, but as a governed intelligence layer that can support real organizational work across communication, information retrieval, process support, and system interaction. Its role is to operate as a trustworthy digital coworker inside an enterprise context.

At the core of NOVARA is Novarion’s Software-defined Brain™, a proprietary architecture designed to unify reasoning, orchestration, contextual processing, and structured tool use into one controlled enterprise AI system. Instead of treating AI as a disconnected collection of point tools, the Software-defined Brain approach turns different capabilities into a centralized cognitive layer for the organization. The objective is not merely to answer prompts, but to make AI usable, consistent, and governable across enterprise workflows.

This matters especially in regulated environments. The more sensitive the organization, the less useful a generic public AI endpoint becomes. Enterprises do not need another detached interface. They need an AI coworker that can be integrated into their own infrastructure, reflect permissions and governance structures, and operate within their own security model.

That is the role NOVARA is built for. It can be deployed as part of an enterprise architecture where data control, integration boundaries, and operational behavior remain in the customer’s hands. It is designed for connection to enterprise systems, knowledge sources, communication channels, and workflow components through controlled interfaces. In practice, that means NOVARA can support activities such as contextual support, internal information handling, structured communication, and orchestration tasks without requiring organizations to export their operational reality into a third-party AI environment.

For regulated organizations, sovereign deployment is therefore not an optional feature around NOVARA. It is a central design principle. If an AI coworker is meant to become productive inside an enterprise, it must be able to run where the enterprise can govern it. That is also why avoiding unnecessary dependency on external cloud API providers matters. The less the intelligence layer depends on external jurisdictions, the more the organization can align AI usage with its own legal, operational, and security requirements.

The infrastructure layer: Xinity Runtime

Xinity Runtime is software that turns an organization’s own servers into an AI inference platform. It installs on existing hardware, behind the existing firewall, and exposes an OpenAI-compatible API on the local network. For an application like NOVARA, this changes exactly one thing: the base URL. Instead of pointing at an external cloud endpoint, the intelligence layer points at a machine in the customer’s own rack. Same API pattern, same request structure, no need to redesign the application architecture around a custom interface.

Underneath that endpoint, Runtime handles what production inference actually requires: deployment and routing of open-weight models, including European sovereign models where the use case calls for them, automatic selection by task and load, audit trails on inference requests, role-based access control, and multi-node orchestration. The core engine is open source under Apache 2.0, allowing customer security and infrastructure teams to inspect the stack rather than accept it as a black box.

For regulated organizations, this is the difference between contractual and architectural control. If prompts, documents, and inference tokens are processed locally, sovereignty is not reduced to a vendor claim. It becomes a technical property of the deployment. No external provider sits in the path by default, and no sensitive enterprise context has to traverse a foreign-operated endpoint in order for the system to work.

Air-gapped deployment is also possible for environments where even internet connectivity is restricted or prohibited. That makes the infrastructure layer relevant not only for commercial enterprises, but also for public-sector, research, and high-security use cases.

In one sentence: NOVARA needs a language model endpoint, and Xinity Runtime makes that endpoint local, inspectable, and aligned with sovereign enterprise operation.

The deployment

In this reference architecture, Xinity Runtime is installed on the customer’s own infrastructure, either on existing servers or on dedicated inference nodes behind the existing firewall. Runtime hosts and serves the selected open-weight models and exposes an OpenAI-compatible endpoint on the local network. From the perspective of the application layer, the integration effort is intentionally small: where a cloud API endpoint would normally be configured, an internal endpoint is used instead.

For NOVARA, this is important because it preserves architectural clarity. The intelligence layer does not have to be rewritten for a special-case on-premise inference engine. The model endpoint remains accessible through a familiar API pattern, which reduces integration effort and accelerates deployment planning.

On the Novarion side, deployment begins with defining the enterprise scope in which NOVARA will operate: which functions it should support, which user groups should access it, which internal systems it should connect to, and which governance requirements apply. From there, NOVARA is integrated into the customer environment as a Virtual Coworker layer with controlled access to relevant enterprise resources. Depending on the scenario, this can include knowledge sources, communication systems, internal workflow components, and permission-aware interfaces.

The implementation approach Novarion typically follows is phased. It starts with a defined use-case scope and a controlled sandbox or pilot domain, then expands into broader enterprise processes as integration, governance, and user acceptance mature. This matters in regulated organizations because deployment success does not come from model access alone. It comes from combining the intelligence layer with enterprise integration, policy alignment, auditability, and operational support.

Around the model endpoint, Xinity Runtime handles routing, load distribution, access control, and traceability. Around the enterprise use case, NOVARA provides the governed coworker behavior, contextual handling, and system-level integration logic. The result is a stack in which the intelligence is useful because it is connected, and the deployment is acceptable because it remains under customer control.

The timeline depends on infrastructure readiness, target use cases, and required integrations. But the key point of this architecture is that sovereign deployment no longer implies a full custom rebuild. The local inference layer and the enterprise coworker layer can meet through a standard interface, while the surrounding implementation is adapted to the customer’s environment.

What this means for regulated organizations

There are three main takeaways from this architecture.

First, AI coworkers do not inherently require external cloud inference. The assumption that capable AI and sovereign deployment are mutually exclusive is no longer technically persuasive. If both the intelligence layer and the inference layer can run inside the customer environment, organizations can pursue useful enterprise AI without surrendering architectural control.

Second, interface compatibility matters. Historically, on-premise AI often meant heavy customization and integration friction. An OpenAI-compatible local endpoint changes that. It allows the application layer to move from external inference to internal inference with minimal disruption, while preserving room for enterprise-specific integration above it.

Third, sovereignty becomes inspectable. That is a major shift for regulated organizations. If no sensitive data has to leave the building for model inference, compliance conversations move away from hypothetical vendor assurances and toward the enterprise’s own infrastructure, controls, and audit processes. That is a much more familiar operating model for banks, healthcare providers, public institutions, and industrial enterprises.

One honest limitation should also be stated clearly: running AI on your own infrastructure means owning its operation. Sovereignty is valuable, but it is not magic. Organizations need the operational maturity to run infrastructure, govern access, select appropriate models, and maintain the environment properly. The benefit is that this responsibility comes with control.

Our point

The important point is not merely that NOVARA can run on Xinity Runtime. It is what that makes possible for regulated enterprises: a credible path toward enterprise AI coworkers that remain useful, governable, and sovereign at the same time.

For Novarion, this architecture shows how the Software-defined Brain™ and the NOVARA Virtual Coworker can be deployed in a way that aligns with enterprise integration and data control requirements. For Xinity, it shows how local, OpenAI-compatible inference infrastructure can turn existing servers into a sovereign AI runtime for serious production use.

If your organization wants to evaluate a sovereign AI coworker architecture for regulated environments, both layers matter: the intelligence layer and the infrastructure layer underneath it.

NOVARA by Novarion brings the enterprise AI coworker layer: trustworthy, integrable, and built for governed use in real organizations. More at https://novarion.ai

Xinity Runtime turns your own servers into a compliant AI inference platform, open source at its core and deployable behind your firewall. More at https://xinity.ai