About Souverana

Advice directly from the person who builds.

Souverana is the AI consultancy of Joel Barmettler, specialised in sovereign AI, RAG and MCP for Swiss companies. No agency overhead, no juniors on the mandate: you work with the person who researches and builds.

Portrait of Joel Barmettler
Joel Barmettler · Souverana, Zurich

Best Paper

ACM Web Conference 2026, GLOW Workshop

35

components in the Swiss AI Hub, air-gapped on one GPU

62

projects delivered as co-founder & CTO

18

engineers led in his own company

Career

  1. 2026 – today

    Founder & Principal AI Consultant

    Souverana · LatentSpace Labs GmbH, Zurich

    Independent consulting for secure and sovereign AI architectures: from Swiss SMEs to international Fortune 500 corporations, across strategy, architecture, systems integration and MCP-based agent tooling.

  2. 2023 – 2026

    Senior AI Consultant & Architect

    bbv Software Services AG

    Architect of the Swiss AI Hub and lead of its five-person core team: an enterprise AI platform of around 35 components, able to run air-gapped on a single GPU or in any Swiss cloud, released under Apache 2.0 and in use in 15 projects for 12 clients. Plus the Swiss AI Agent Protocol with three SDKs, and the build-up of a new AI business unit including two public tenders.

  3. 2019 – 2023

    Co-founder & CTO

    PolygonSoftware, Zurich

    Co-founded at 22 while studying and grown from 3 to 18 engineers in four years: 62 projects for 25 clients in five industries, including a computer-vision system for parcel recognition for Swiss Post.

  4. 2019

    AI Engineer

    BE Semiconductor Industries (Besi), Zug

    CNN-based defect detection for die bonding in semiconductor manufacturing: 99 percent accuracy at several thousand chips per hour, integrated into the running production line.

Research & Open Source

ConceptFormer: Towards Graph-Native Grounding of Large Language Models via Latent Concept Injection

ACM Web Conference 2026 (WWW ’26), GLOW Workshop · Best Paper Award

Grounds language models directly in knowledge graphs without a text detour and improves fact-retrieval accuracy by 348 percent. With Abraham Bernstein and Luca Rossetto, University of Zurich.

Progressive in Principle, Centrist in Practice: LLM Political Bias Is Instrument-Dependent

arXiv preprint, 2026

Shows that the political bias measured in language models depends on the measuring instrument. They answer left of centre on abstract questionnaires, while on real Swiss referenda they sit closer to Die Mitte and the FDP. Based on 75 Smartvote questions put to 66 models and 48 federal referenda.

arxiv.org/abs/2606.00048

Open source: the Swiss AI Hub (Apache 2.0) plus the publicly usable tools gheim (PII protection for LLM access), vue-mcp and slidev-mcp.

Education

MSc in Computer Science, University of Zurich

summa cum laude · 2024

Major in Artificial Intelligence (grade 5.9). Master’s thesis "ConceptFormer" awarded the top grade of 6.0.

BSc in Computer Science, University of Zurich

summa cum laude · 2021

Software Systems, minor in Neuroinformatics. Bachelor’s thesis awarded the top grade of 6.0, implemented in the open-source rendering engine appleseed.

Talks & Media

  • 2026

    CH Open Source AI Conference

    Workshop: a RAG system in six hours, built with open-source coding agents.

  • 2024

    Swiss AI Conference, Rotkreuz

    Hands-on workshop on AI agents on the Swiss AI Hub.

  • 2024

    Swiss AI Impact Forum

    Panel on the technology, ethics and economics of AI, with SwissCognitive, ZHAW and the ETH AI Center.

  • 2023 – 2026

    Keynote & webinars

    Opening keynote on generative AI and six public webinars, from ChatGPT internals to AI agents.

  • 2023 – 2026

    "Attention Heads"

    German-language AI podcast, 27 episodes: technical deep dives, industry analyses and societal questions.

The Approach

The distance between what AI can do in a paper and what it delivers in production for paying users is where Souverana works. That takes research and engineering, plus respect for the difference between a demo and a system that holds.

Anti-hype is not a style here but a method: for regulated, data-sensitive Swiss companies, what counts is not the biggest model but control, traceability and a result the data protection officer can sign off on.

LinkedIn · GitHub · ORCID · gheim.ch