Език: English
AI sovereignty is often reduced to the question of whether a model can run locally. But the model is only one component of a much larger system.
This talk looks at sovereign AI as an architectural problem. Using PrAIvate, a privacy-focused RAG platform built exclusively with free software apart from the models themselves, it examines how retrieval, document processing, inference, access control, APIs, deployment, and user interaction combine into a complete AI system.
The central idea is that sovereignty means preserving meaningful choices: the ability to inspect, replace, migrate, integrate, and operate the components of the system independently.
Organizations often start their AI journey by asking which model they should use. Yet models are changing rapidly and are becoming increasingly interchangeable. The architectural decisions around them tend to be much more persistent.
This talk presents a different perspective: the model is not the system. AI capability emerges from the interaction of many components — document ingestion, retrieval, embeddings, reranking, inference, identity and access management, APIs, user interfaces, deployment, and operations.
PrAIvate serves as a practical example. It is a privacy-focused RAG platform whose software stack, apart from the models themselves, is built exclusively with free software. The architecture is designed to avoid unnecessary dependencies on individual vendors, cloud platforms, or model providers.
A key requirement is deployment independence. The same system should be able to run in very different environments: locally on a personal computer, on-premises in an organization’s own data center, or in cloud infrastructure. The deployment model should be a choice, not an architectural constraint.
The talk will discuss several principles behind this approach:
separating the AI model from the surrounding system architecture
designing components to be replaceable rather than tightly coupled
combining different retrieval methods instead of relying on a single mechanism
keeping organizational data and access control under explicit control
using free software to preserve inspectability, portability, and operational independence
designing for deployment across local, data-center, and cloud environments
The goal is not technological isolation and not the claim that every organization should operate every component itself. Sovereignty means retaining the ability to make meaningful choices as requirements, models, and infrastructure change.
The broader argument is that sustainable AI depends less on choosing the “best” model today and more on building systems that can survive the models, vendors, and platforms of tomorrow.