Home Server 2 Electric Boogaloo

There have been some updates to the home server, I thought it might be helpful to write about them.

Home Server

It now has a battery. This has been helpful to prevent any drops in power, I’m also using tlp to dictate the battery to always hover around 50%.

It is also 14 TB now! I appended a 14TB external drive, and am backing it up as well on a separate drive.

It’s also using Tailscale now, instead of a WireGuard on the local network, because of my ISP introducing CGNAT (Carrier-Grade NAT). This means my external IP is shared across others in the area, breaking port forwarding. Tailscale circumvents this issue.

Nexus

I’ve introduced a simple web app which links to other services that are handy to get to. Immich and Radicale are still there, like previously. I’ve re-done the Beast toggle to be here, turning it on again uses Wake On LAN. Turning it off is a system service on the Beast.

LibreChat, Journal Transcriber, and Recording Transcriber are all new, and hosted on the beast within the Tailscale network.

Beast

Specs

The main new thing hardware-wise about the beast is an RTX 5060 Ti 16GB VRAM. This allows for some better models than previously.

LibreChat

LM Studio is used to run LLMs now. I can tweak context size, run different kinds of models and I generally find it more useable than Ollama. The main LLM is Gemma 4 12B at the moment.

On startup (without logging in), LibreChat is started using Docker. LM Studio has a daemon that can run on startup as well, so everything is ready to go as soon as the Wake On LAN request is sent from the home server. I’m finding this to be mainly helpful for long documents, sensitive queries, that kind of thing.

Journal Transcriber

I like to write some things by hand, like complex ideas, or diary entries. I got GLM-OCR running using LM Studio, and this web app queries that API with uploaded images. For example here’s a short entry.

…And then transcribed.

Recordings Transcriber

The recordings transcriber runs two models for each video uploaded.

1. NVIDIA Parakeet: Offers very fast transcription of recordings.

2. Pyannote: Used for speaker diarisation, very helpful when then feeding it into librechat to extract insights.

It outputs as VTT as well, so I can then watch the recording back with subtitles.

Shutdown

The shutdown script is more or less the same as described previously, except login is no longer required. It now runs as a system-wide background service.