The AI stack runs on a dedicated Ubuntu VM (ai01) with full GPU passthrough. Docker Compose orchestrates all services: Ollama handles model inference with a curated lineup (gpt-oss:20b, mistral-small3.2:24b, qwen3:14b, qwen3-coder:30b), Open WebUI provides the chat interface, Qdrant powers vector search, and OpenClaw enables multi-agent workflows with AWS Bedrock integration. Models are managed via Docker bind mounts for persistence across container rebuilds.