I build AI agents that can be trusted with real work.

I am a Senior AI Engineer in Zürich. My systems automate customer operations: agents that read, route, and resolve the requests that used to fill a team’s inbox.

What separates a production agent from a demo is rarely the model. It is the machinery around it: guardrails that bound what an agent is allowed to do, evaluation that shows whether it actually helps, and data foundations that both depend on.

Above the individual agent sits the layer I find most consequential, the agent mesh: the control plane that lets an organization run many agents and still keep a human behind the steering wheel.

Selected work & credentials

• Peer-reviewed: Practical Strategies for Enhancing Reliability of GenAI Systems in Customer Operations — SwissText 2024, ACL Anthology.
• Speaking: Invited speaker at the SwissNLP "LLMs in Practice" expert group. Workshop speaker at the CH Open Workshop-Tage.
• In production: I build and run these systems in production, not on slides.
• Upcoming: On 10 September 2026 I teach a full day workshop at the CH Open Workshop-Tage in Rapperswil: Enterprise-Ready AI Agents: Architecting the Agent Mesh with Open Source. Tickets for the day are available on Eventfrog.

Selected writing

• Start here The Agent Mesh May 2026 The control plane for the agent era: how an organization runs many agents and still keeps a human behind the steering wheel.
• A short history of looking June 2026 Why the only prompting technique that does not go stale is understanding what the model attends to.
• Guardrails for AI Agents September 2025 A risk-based approach to AI agent safety.
• A Domain-Driven Approach to MCP September 2025 Why business domains, not technology stacks, are the right foundation for dependable agents.

Newsletter

I write about this work in my newsletter, Undisconnected: guardrails, evaluation, and the control planes that keep a human behind the steering wheel. If you are building agents that have to survive contact with production, it is written for you.

I take on a small number of advisory conversations and speaking engagements. The fastest way to reach me is LinkedIn.

Elsewhere

My professional work is proprietary, but you can find open-source experiments on GitHub and connect with me on LinkedIn.