Homelab + AI + Self-hosting
Tinkering with servers, running local AI models, and connecting it all together. I self-host everything I can and document the journey here.
Say hi01 What I do
Mostly servers, AI setups, and whatever rabbit hole I fall into next.
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.
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.
Web hosting, email, DNS, git repos, CI runners. If there's a self-hosted alternative, I'm probably running it somewhere.
Pipelines, deployment scripts, and AI agents that do the boring stuff so I don't have to. Still tweaking, always tweaking.
Encrypted tunnels everywhere, key-only SSH, firewall rules, monitoring. Not because I'm paranoid, but because getting hacked would be really annoying.
02 Projects
Side projects, tools I built for myself, and things that started as "this should be quick" and never were.
03 The Stack
A mix of dedicated servers, mini PCs, and a GPU workstation, all stitched together with Tailscale.
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.
An RTX 3090 running Ollama and ComfyUI, serving local AI to every machine on the network. No cloud, no API keys, no monthly bills.
Tailscale mesh VPN connecting everything. Home, data center, laptop, phone. Every connection encrypted, every device reachable by name.
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
Write-ups about things I've built, broken, and eventually fixed.
Four quant tiers of a mixture-of-experts model, a dense 27B with MTP and a 321B MoE with CPU offload on two RTX 3090: 2 bits run 1.53 times as fast, cost five of fifty exact answers and tip over in agentic use.
A kernel module frees the halved math rate of GeForce cards. I measured it: the rate per clock cycle really is twice as high — my language model only gets 7.5 percent faster. I wrote the probe that makes the difference visible myself.
The same model, the same 220 probes, the same judge, temperature zero: on llama.cpp it answered 108 of 110 adversarial probes, on the new engine only 100. Over-refusal in both runs: not one of the 110 neutral control questions denied. What changed was the engine — and the draft layer thinking along. The protocol and the scorecard are attached as files.
05 Contact
Based in Europe. Always happy to chat about homelab stuff, AI, or whatever you're building.