PLAIN: A Sovereign Data and AI Platform for Germany

Germany's federal ministries were sitting on data they couldn't use together. Each department ran its own tools, its own pipelines and its own vendor contracts. When a question spanned departments, whether pandemic response, climate policy or supply-chain risk, the analysis took months of coordination, if it happened at all. The usual shortcut was off the table, because sovereignty requirements ruled out the hyperscalers, and German public procurement sets its own pace.

The brief for PLAIN (Platform Analysis and Information Systems) was one shared data and AI platform for five federal ministries, run from a sovereign Berlin data centre and operated by Bundesdruckerei under the Federal Foreign Office. It had to be fully accessible under BITV 2.0, free of vendor lock-in, and robust enough for everything from crisis monitoring to funding allocation.

Earlier in my career I treated compliance the way most teams do: build the product, then work out how to make it pass. That ordering reliably produces late structural surprises. An end-of-cycle review finds foundational problems, because the architecture settled months before the rules entered the room.

On PLAIN I inverted it. KRITIS, BSI controls, GDPR and BITV 2.0 shaped the first architecture decision. A federal requirement can't be argued down in a workshop, so a lot of debate never started, and the architecture came out simpler for it. The trade was accepting restrictive defaults up front: sovereign infrastructure only, isolation everywhere, everything auditable. In exchange, the regulator never met us as an adversary at launch.

A containerised multi-tenant architecture with network and identity isolation per ministry. A data layer combining lake, warehouse, data contracts, lineage and catalog. Self-service workbenches matched to the roles people held: Apache Superset for no-code dashboards, GitLab for inner-source collaboration, Jupyter and ML toolchains for data science teams. Virtualised GPUs with quota policies kept AI workloads elastic without runaway costs.

Adoption was part of the delivery work, because infrastructure nobody uses helps no ministry. Analysts, data scientists and policy advisors each got a workflow shaped for them: governed data onboarding with PII handling, dashboards mapped to ministerial KPIs, reusable container templates that cut use-case rollout from weeks to days. Sandbox tenants with synthetic datasets gave teams room to experiment without touching production data. Cross-ministry sharing ran on inner-source patterns, scoped tokens and data-sharing agreements, so a ministry could reuse another's work and keep control of its own data.

Version 1.0 reached production in June 2023, roughly eight months after development started. The platform took 2nd place in the eGovernment Competition for "Digital Transformation through AI and Modern Infrastructure."

What runs on it today: pandemic and political crisis monitoring with scenario insights, climate-aware land management planning, supply-chain criticality forecasting, optimised funding programs, BMZ data products with AI-assisted dashboards. Two years earlier, none of it existed.