Workshop: Kontrolle statt Chaos - Aufbau einer souveränen Agentic AI Platform

This is the companion repository for a workshop at the Cloudland conference in Soltau on May 21st, 2026.

Abstract

Companies increasingly recognize the strategic importance of Artificial Intelligence, yet deploying AI solutions at scale presents significant challenges. This hands-on workshop provides a practical, step-by-step guide to designing and building a cloud-native AI platform.

You'll learn how to create a unified platform that simplifies the entire AI lifecycle using Kubernetes.

What you'll build:

• Agent runtime and orchestration layer
• API gateways for unified agent and model access
• Observability and monitoring stack

This repo guides you through the process of deploying agents onto a pre-provisioned AI platform.

Overview

The cluster is shared across all participants. The platform (operators, gateways, monitoring, LibreChat) is already deployed. You get your own namespace (ns-XX) on the shared cluster to deploy your agents into.

• Step 0: Tour the pre-deployed platform
• Step 1: Deploy your first agents (and chat with them)
• Step 2: Evaluate your agent with an Experiment
• Step 3: Protect your prompts with a PII guardrail (Presidio)
• Step 4: (Bonus) GitOps + alternative UIs (Flowise, Langflow)

---

Setup

1. Fork this repository.

2. Create a GitHub Codespace on your fork.

3. Claim a namespace by entering your name (or just an emoji) next to one of the ns-XX slots on this spreadsheet: https://ethercalc.net/mzku2p0zjwkk

4. Decrypt the shared kubeconfig (Kubernetes credentials file). The password will be announced during the workshop.

   ./decrypt-kubeconfig.sh kubeconfigs/workshop-kubeconfig.yaml.enc <password> kubeconfig.yaml

5. Connect to the cluster. Replace ns-XX with the namespace you claimed.

   export KUBECONFIG=kubeconfig.yaml
   export YOUR_NAMESPACE=ns-XX
   
   kubectl get pods -n $YOUR_NAMESPACE   # should be empty
   kubectl get namespace $YOUR_NAMESPACE  # confirms your namespace exists

6. Start hacking! Continue with Step 0.

---

Port-forward tip

Every step that uses LibreChat or Grafana asks you to kubectl port-forward. The connection dies on every server-sent-events close, so run it in a retry loop (and force websockets, which are more tolerant of idle gaps):

while true; do
  KUBECTL_PORT_FORWARD_WEBSOCKETS=true \
    kubectl port-forward -n librechat svc/librechat-librechat 3080:3080
  sleep 1
done

Swap the namespace + service to forward Grafana (monitoring/grafana 3000:80) or the observability dashboard (observability-dashboard 8100:8000).

---