I consult under the name Souverana, independent AI consulting and architecture, based in Zurich. The legal entity behind it is LatentSpace Labs GmbH, which I founded in July 2026 after building the enterprise AI business area and the Swiss AI Hub at bbv Software Services.
The premise is short: your data is confidential, so your AI has to be too. Administrations, hospitals, and law firms can run AI on real, confidential data today, in Switzerland and under their own control. The technology is not the bottleneck. What is missing is someone who assembles the available pieces so they hold for a specific organisation's data.
Switching on Microsoft Copilot and hoping the rest follows. Handing developers GitHub Copilot licences and waiting for productivity. Naming someone "AI Engineer" and having them assemble a RAG prototype.
If that describes your organisation, you are in good company. A strategy starts somewhere else: where AI creates value in your business, which data it needs to do that, and how it reaches production safely.
It is not a product. It is four layers, and each one has good open options today.
Above the four layers sits the disposition. Organisations that get somewhere with AI try things early and get used to change, and that is part of what I bring to an engagement.
Shadow AI already shows where the demand is. Staff are using ChatGPT with case files and patient data, which means the task is real and the tool works. Those are the use cases waiting for a solution, and the first step is an official route to them.
The components exist: open models such as Apertus, hosting in Switzerland or in-house, mature tooling for RAG and MCP. The EU AI Act and the revised Swiss Data Protection Act belong in the build plan from the start, carried along as construction details so the architecture satisfies them by construction. Bolting compliance on at the end is what makes it expensive.
Independence is what makes the recommendation worth anything. Hyperscalers, platform vendors, and large consultancies all arrive at their own product. I sell neither a platform nor a model, so the advice can follow your situation instead.
You start small and with a clear frame, and each step produces something the next one builds on: an assessment of where you stand and what is exposed, then the sovereign target architecture across all four layers with model, hosting, and platform chosen independently, then the build and the coordination of hosting and operations partners, then ongoing work as your AI architect through operation, rollout, and further development.
Most AI consulting ends with a slide deck. That is roughly where the interesting problems start: how documents actually get ingested, what happens when the model is wrong, who may see which retrieval result, how a failure is traced six months later. So the concept is where the work begins. I build, and I stay, including as a fractional AI lead where that fits better than a project.
Rates are published rather than quoted on request: fixed prices for the assessment and the target architecture, a day rate for implementation, a monthly retainer for the fractional lead role. The initial conversation and the sovereignty check cost nothing.
Organisations of every size, from sole proprietorships to international corporations. Three mandates are running, spread across the four layers.
Integration. A Fortune 500 company (name confidential) is making its own software accessible to AI through MCP servers. I am building part of an environment in which models can reach a large body of data, reason over it, and act on it.
Platform and infrastructure. I continue to co-own the architecture of the Swiss AI Hub, the Apache 2.0 platform I originally built as AI Architect at bbv. A complete AI stack, deployable with one command, air-gapped on a single GPU or in any Swiss cloud. It solves precisely the infrastructure and platform problem described above, and it is the reference implementation for much of what I recommend: clients can read the code.
Strategy. A Swiss SME (name confidential) on its internal AI strategy, focused on AI in software engineering and on a sovereign AI transformation.
Website: souverana.ch · Sovereignty check: souverana.ch/check
Souverana is the independent AI consultancy Joel Barmettler runs from Zurich for organisations working with confidential data. It covers sovereign AI across four layers, strategy, infrastructure, platform, and integration, and continues past the concept into implementation and operation. Clients range from sole proprietorships to international Fortune 500 corporations.
Souverana is the consulting brand. LatentSpace Labs GmbH, founded in Zurich in July 2026, is the legal entity behind it and remains the operator of joelbarmettler.xyz.
Strategy, where AI creates value in the business and in what order to proceed. Infrastructure, the language model and its hosting, vector database, document storage, and data and knowledge pipelines. Platform, the tools the team works with day to day, wired into a coherent whole. Integration, the agents, RAG, and MCP connections into existing systems. Each layer has good open options today; the work is selecting them and assembling them so they hold for a specific organisation's data.
Yes, as long as the processing stays under the organisation's control. The revised Swiss Data Protection Act (revDSG) does not prohibit AI, it requires control, transparency, and proportionality. With processing in Switzerland or on-premise, clear authorisation concepts, and traceable sources, AI can be justified even under professional and official secrecy.
A stack of open components: a language model such as the Swiss Apertus, hosted in Switzerland or in-house, with a platform and an integration layer the organisation controls. For most enterprise tasks, knowledge search, document analysis, and assistance, this reaches production-grade quality.
Both. The architecture and the integration layer, meaning the agents, RAG, and MCP connections, are built in-house. Hosting and operations are handled by established Swiss partners, coordinated by Souverana, so the client keeps a single point of contact through to operation.
.