Everything AI-related in the lab and where it runs. Ground rule: models run on our own hardware — no hosted-model APIs and no free-tier API sign-ups.
| Project | What it is | Where | Pages |
|---|---|---|---|
| Jarvis | Private assistant: chat over ntfy topic jarvis, permission tiers, audit log, git-versioned markdown notebook. Reads Proxmox/Ceph/backups, Home Assistant, the wiki and the cameras, with opt-in proactive alerts; with Cody's yes it can start/stop/restart containers, edit the backup job and run a command in a container. Local Gemma 4 26B-A4B (CPU). |
LXC 130 on SuperPC | Plan and progress log · Service · User Guide |
| Homelab Monitoring Agent | n8n workflow (every 30 min) that classifies cluster health into three tiers with the local LLM and can run a whitelisted set of restarts through action-executor. |
n8n LXC 105 · Ollama LXC 126 · action-executor LXC 127 | Agent Hub (history) · n8n |
| Shared Ollama | One CPU Ollama backend for the monitoring agent and Odysseus. | LXC 126 | Agent Hub |
| Odysseus | Self-hosted AI research/chat app. | LXC 111 (stopped as of 2026-10-09) | Odysseus |
| Terraria AI bot | Bot client for a Terraria server. | LXC 128 (stopped as of 2026-10-09) | no page yet |
All inference is CPU-only on SuperPC (Ryzen 9 9900X, 60 GiB RAM, no usable GPU). Rough numbers measured for Jarvis: ~25 tokens/s generation, ~185 tokens/s prompt processing — prompt processing is the real bottleneck. A GPU (RX 580 in the open x4 slot) is under consideration; see SuperPC and the Jarvis plan. Speech-to-text (Whisper) and text-to-speech (Piper) are installed for the voice plan and measured, but not enabled.
What went wrong and what fixed it while building these: Lessons Learned.