IT administrator by day (document management and digitalisation in public administration), AI builder by night. Trained as an IT systems specialist, studied IT security, taught myself the rest. I run my own AI stack at home in Balingen, Germany — one workstation, a small k3s cluster and a lot of nights — and I publish what I find along the way.
No datacenter. No enterprise budget. Just consumer and workstation hardware and a stubborn refusal to accept "unsupported" as the end of the story.
FLE — Fast Local Experts · v1.1.0 · MIT A llama.cpp fork for running large Mixture-of-Experts models on one 32 GB GPU with the routed experts in host RAM: a GPU expert cache with a byte budget and warm start, cache hits on the GPU overlapped with misses on the CPU, MTP speculative decoding with the model's own head, and a leaner prefill. On Qwen3.8-Flash-Next (89 GB) it decodes at roughly 90 tok/s in my own paired measurements (RTX PRO 4500 32 GB + Ryzen 9950X3D). Technical report · Zenodo · GGUF models · REAP448 · “Making Rhea fast” films
FLE Studio · v0.3.0 · MIT A local chat and lab for llama.cpp and the FLE engine: GGUF model loader, live inference data, launch profiles. No account, no telemetry, macOS, Linux and Windows. Website
FYOS — For Your Own Sake · v0.1.1 · Apache-2.0 A personal operating system in the browser that runs entirely on your side: a desktop shell with windows and apps, an MCP hub that bundles tools and interfaces from plugins, an installer for Docker Compose or Kubernetes, and an agent that works with those tools and asks before heavy steps. fyos-os (shell, hub, installer) · fyos-agent (the agent, an OpenClaw mod) · fyos-plugin-vorlage (template for your own plugin)
Contexierge · Claude Code plugin · MIT Claude keeps its own session notebook, injected verbatim after every compaction, so what matters survives long sessions instead of depending on the automatic summary.
| Project | What it is |
|---|---|
| consumer-llm-patches | 11 patches from fine-tuning Gemma4 26B on a single RTX 4090, with an honest per-patch validation status |
| three-level-ai-development | A method for building with AI: human vision, persistent architect, execution AI |
| snowmelting-method | Companion method to the above |
- Every claim gets a measurement, and I retract when new evidence says otherwise.
- I build with several AI coding agents in parallel; the decisions stay with me.
- I build for myself first. Sensible issues, PRs and hints are welcome; there is no support promise.
🌐 sirfyyn.de · 🤗 huggingface.co/sirfyyn ·