Hands-on, on your stack, for the people who have to run it on Monday.

Workshop topics

What we actually do in four hours.

Running AI infrastructure: Docker and GPUs

From a single compose file to a GPU server that serves open-weight models: containers, the inference stack, Traefik in front, backups that provably restore, upgrades, and the runbook for the middle of the night. The material behind the Udemy course, pointed at what an AI server needs in 2026.

Claude Code in a real workflow

Not a demo. We take one task your team does weekly and run it through Claude Code end to end: the CLI, project instructions, tool use, an MCP server that talks to one of your own systems, and the review discipline that keeps it safe. You leave with a working setup in your repo.

Agents with a role, a context and guardrails

Purpose-built agents — one for intake, one for operations, one for review — each with a defined role, its own context, and boundaries for the untrusted data it reads. We use OpenClaw where a multi-agent setup fits, and plain scripts where it doesn't. The lesson is the design, not the framework.

• Model routing that uses a small model for triage and a large one for reasoning — the usual result is a 60–80 % cost cut with no loss in output quality
• Handling emails, feeds and documents you did not write, without letting them steer the agent
• Connecting to the tools you already have: mail, calendar, databases, your own APIs
• What to log so that someone can audit what the agent did, and why

Swiss data rules for the people who build

What may go into which tool under nFADP, where the CLOUD Act bites, and how to split work between a Swiss-hosted model and a frontier model without guessing. This is the practical version of the AI Exit on the offer page, taught to the team instead of built for them.