Building AI agent fleets on self-hosted hardware since before it was cool.
My public repos are the visible edge of a homelab AI fleet — multiple servers, dozens of agent profiles, local LLM serving, and memory systems that learn while I sleep. The philosophy: own your inference, measure everything, automate the maintenance, and publish what it teaches you.
The fleet's daily-driver machines run Omarchy (Arch + Hyprland), and when the desktop is missing something, I build the plugin and publish it.
| Repo | What it is |
|---|---|
| praxis | ⭐ The hub — hard-won lessons from running the fleet: serving recipes with verification gates, agent memory architecture, benchmark methodology, curated links |
| mnemograph | The in-process knowledge graph for AI agents — no external databases, no network calls, zero infrastructure |
| focus | Self-hosted web launchpad — Flask + vanilla JS, the start page every browser in the fleet calls home. Inspired by Bonjourr. |
| Clearvoice | Privacy-first meeting assistant for Linux — record, transcribe, diarize, and summarize meetings with local AI. The fleet's meeting-intelligence frontend. |
| bazzite-customizations | KDE Plasma customizations and plugins for Bazzite (immutable Fedora) — sudo-free, reversible, measured. First entry: push-to-talk dictation via remote WhisperX. |
| omarchy-recipes | The Omarchy twin of praxis — how the fleet's laptops run a heavily customized desktop as code: recipe-repo discipline, measured update-overwrite model, plugin hygiene |
Five open plugins that grew out of running Omarchy on the fleet's laptops:
| Repo | What it does |
|---|---|
| omarchy-appdock | KDE-style task manager and minimize engine for Omarchy (Hyprland 0.56) — dock widget, per-workspace window icons, hyprbars titlebar buttons |
| omarchy-tailfin | The stock Tailscale bar widget rebuilt as a tabbed panel — health warnings, preference toggles, exit-node and Mullvad tabs, searchable peer list |
| omarchy-sysmon | Live RAM, CPU, network rate, and whole-disk stats in the Omarchy bar — four groups, one widget |
| omarchy-displayplus | One panel for every display setting — brightness, text size, and universal scale on the front page, full hyprmoncfg layout editor underneath |
| omarchy-powercore | Battery and power strategy — per-source profiles, charge protection, clamshell inhibitor, per-source keyboard backlight, live watts |
- praxis — the operating education: what deploying all of this taught us
- mnemograph — knowledge-graph memory as an embeddable library
- Clearvoice — meetings in, structured notes out, without the audio leaving the machine
- Omarchy plugins — the desktop layer the agents and I both work from
- omarchy-recipes — how those Omarchy machines are run as code
- bazzite-customizations — the KDE/Bazzite twin: Plasma tweaks and plugins for the immutable-Fedora desktops
- Everything else (startpages, tooling, forks I track) keeps the fleet running
- Qwen3.8-Flash-Next-NVFP4-DGX-Spark — vLLM serving of Qwen3.8-Flash-Next (NVFP4) on DGX Spark: disk-backed n-gram table patch, staged gather, reduced-vocab MTP draft. 43.9 tok/s on one Spark, 53.7 tok/s TP2, 9M-token KV pool at TP4. Launchers, patch, and a 40-prompt harness included.
- hermes-agent — the agent framework the whole fleet runs on
- BonjourrNue — minimalist startpage (upstream fork I track)
- build.nvidia.com — where the DGX Spark practitioner community trades notes (find my verification posts there)