Descriptions:
Bilge Yücel, senior developer relations engineer at deepset GmbH — the company behind the open-source Haystack orchestration framework — walks through the practical engineering consequences of building AI systems under sovereignty constraints. The talk is grounded in deepset’s work with European enterprise clients including Airbus, Bosch, and Siemens, as well as public sector organizations like the European Commission and Germany’s Federal Ministry of Research and Technology.
Yücel structures sovereign AI across four pillars: data sovereignty (GDPR compliance requires that EU citizen data not be processed by US-hosted APIs like embedding services running in Virginia), model sovereignty (avoiding tight coupling to a single provider), infrastructure sovereignty (the spectrum from air-gapped deployments to SaaS and the Cloud Act risk at the far end), and operational sovereignty (auditability, version control, and incident ownership). Each pillar is mapped to specific compliance regimes and the failure mode that occurs when it is violated.
The second half of the talk focuses on what actually breaks when organizations try to retrofit sovereignty onto existing systems: prompt and pipeline rewrites when switching models, search fragmentation across multiple jurisdictions, sudden Kubernetes management burden after moving off managed infrastructure, GPU-CPU network connection complexity, and the discovery of fully opaque AI application layers with no observability. Yücel presents Haystack’s consistent component interface, typed data flows, and YAML-serializable pipeline format as engineering solutions to several of these problems. This is a practical reference for teams building RAG pipelines or agent systems in regulated European or public sector environments.
📺 Source: AI Engineer · Published May 19, 2026
🏷️ Format: Deep Dive
