Homelab + AI + Self-hosting

I build things and write about it

Tinkering with servers, running local AI models, and connecting it all together. I self-host everything I can and document the journey here.

Say hi

01 What I do

Things I enjoy building

Mostly servers, AI setups, and whatever rabbit hole I fall into next.

Homelab & Servers

Proxmox clusters, LXC containers, VMs for everything. I like building infrastructure that just runs, even when I forget about it for weeks. Multiple nodes at home and in data centers, all connected through a mesh VPN.

Local AI

Running LLMs and image generation on my own hardware. No cloud APIs, no per-token billing. Just a GPU doing its thing on a shelf in my room.

Self-hosted Everything

Web hosting, email, DNS, git repos, CI runners. If there's a self-hosted alternative, I'm probably running it somewhere.

Automation

Pipelines, deployment scripts, and AI agents that do the boring stuff so I don't have to. Still tweaking, always tweaking.

Security

Encrypted tunnels everywhere, key-only SSH, firewall rules, monitoring. Not because I'm paranoid, but because getting hacked would be really annoying.

10+

Years tinkering

Multi

Server locations

100%

Self-hosted

24/7

Always running

02 Projects

Stuff I'm working on

Side projects, tools I built for myself, and things that started as "this should be quick" and never were.

ClawForge An AI-powered server management tool. Think hosting panel, but the AI does the work instead of you clicking buttons.
AI Homelab Local LLM inference and image generation running on my own GPU. All my servers can use it through a mesh VPN.
Web Hosting Running sites and services for myself and a few others. Mostly WordPress, some custom apps, all self-managed.
Home Automation Monitoring dashboards, automated backups, alerting. Making the infrastructure take care of itself.
Open Source Contributing back when I can. Building tools I wish existed, sharing configs that took forever to get right.

03 The Stack

What's running under the hood

A mix of dedicated servers, mini PCs, and a GPU workstation, all stitched together with Tailscale.

Compute

Proxmox VE clusters with a bunch of VMs and containers. Dedicated servers in data centers plus a few machines at home. Overkill for one person, but that's kind of the point.

AI Hardware

An RTX 3090 running Ollama and ComfyUI, serving local AI to every machine on the network. No cloud, no API keys, no monthly bills.

Networking

Tailscale mesh VPN connecting everything. Home, data center, laptop, phone. Every connection encrypted, every device reachable by name.

Monitoring

Dashboards, alerting, automated backups. I get a notification before things break, and snapshots so I can roll back when I inevitably break something myself.

04 Latest

From the blog

Write-ups about things I've built, broken, and eventually fixed.

Bar chart with five groups on a very dark ground, one violet bar for adversarial probes answered out of 110 and one green bar for neutral control questions refused out of 110. Stock Q4_0, marked as the control: 12 answered, 16 controls refused. Abliteration Q4_K_M unsloth: 110 answered, no green bar. Abliteration Q4_K_M as fleet default: 110 and no green bar. Abliteration Q2_K: 107 answered, no green bar. GB10 Flash-Next in IQ3_S: 108 answered, no green bar.

Oct 22, 2026

One switch, 89 percentage points: what a guardrail test shows about uncensored models

A guardrail test asks not whether a model answers well but whether it answers at all. Five builds of one family: the stock build answered 12 of 110 adversarial probes and refused 16 neutral questions, the four abliterated builds answered 107 to 110 of 110 - and refused not one control question.

Diagram of the request paths in the local LLM fleet: clients (pi, llm-local, Claude Code, Paperclip, 25+ hosts) send chat to the Atlas llm-router on port 11434; qwen3.8:27b requests are forwarded to the GPU llama.cpp server with six 128k slots at up to 109 tokens per second aggregate, embeddings and utility models go to the internal ollama on 11436, prompts over 128k tokens go directly to the GB10 DGX Spark with four 262k slots, and the laptop with an RTX 2080 Ti serves short interactive tasks locally at 54.7 tokens per second.

Oct 20, 2026

Three machines, one endpoint: how the local LLM fleet actually runs

A dual-3090 server with six 128k slots, a DGX Spark for everything over 128k, a laptop faster than both, and a 200-line router that keeps every client on the old port.

Two-field panel on a very dark ground. Left, the old setup: llama-server with -c 262144, -np 4 and a unified KV cache, where each of four slots may grow to 262k tokens but all four share 262,144 cells, so an overflow aborts every running request. Right, the fix: --kv-unified-per-slot 262144 with a pool of 1,048,576 cells stored as q8_0, so no slot can starve another. The strip below lists the measured price: 13.2 GiB more memory instead of the estimated 6.8, 24 failed restarts and a 7-minute outage caused by an unquoted value in the environment file, and q8_0 speed not yet measured. Footer marks: 967 errors in 3 days, 4 slots, 262,863 greater than 262,144.

Oct 17, 2026

Four agents, one cache: why my LLM server killed every session at once

My coding agent kept dying with Error: Context size has been exceeded. Its status line said 16.7 % of 262k, so about 44,000 tokens. Auto-compaction was switched on. By…

All posts →

05 Contact

Get in touch

Based in Europe. Always happy to chat about homelab stuff, AI, or whatever you're building.

tobias@liauw-media.com