Cecilie Monsen

How do you deliver modern, agentic AI services when your systems and data are subject to the uncompromising frameworks of the Norwegian Security Act (SSA)? At Telenor, we faced this exact dilemma: the need to build an on-premise, air-gapped, highly secure AI environment with absolutely zero margin for error.

In this session, we introduce "Pythia", our production-grade, hybrid AI cluster. We will explore the technical realities of combining virtual machines with HPE bare-metal GPU nodes to create a robust architecture, while navigating the issues surrounding a multi security class service.

The core of our strategy relies on Zero Trust principles, bypassing massive classification hurdles by treating the central AI cluster strictly as a stateless data processor. By enforcing strict session isolation and keeping all access to classified data at the client level we managed to thread this through the needles eye.

Attendees will learn how to:

• Implement a stateless AI architecture designed to handle multiple security classes simultaneously.
• Design remote RAG architectures across distinct security boundaries, keeping classified databases and API endpoints strictly separated from the AI processing cluster.
• Logging and telemetry when handling classified data

Join us to discover how to push the boundaries of enterprise AI and open-source flexibility without compromising strict national security compliance.