Hey, I'm Fred

AI Infrastructure Architect. I build on Omarchy, with an AI that lives on it.

Two obsessions, and they turned out to be the same one.

Omarchy — DHH's opinionated Arch + Hyprland desktop — is what I run every day. I build plugins, forks and patches on top of it.

PAI is the Personal AI Infrastructure I built around Claude Code: skills, hooks, memory, and a repeatable algorithm, so an AI collaborator holds context across months instead of minutes.

They meet on one screen. The AI writes the plugins. The plugins report on the AI. Memory keeps the score so nothing gets built twice.

Omarchy ships a beautiful bar. I kept adding things to it until the bar became my operating surface: messages, cars, backups, Bluetooth, media, agents, weather, notifications. Every one of them is a real repo, MIT-ish, installable, theme-driven.

Here it is under load — music playing, a local model on the GPU, Codex mid-session:

Other people's plugins, credited to their authors. I keep forks so I can patch them for my machine and send the fixes back upstream:

Also live on my bar but not yet cut loose as public repos: Burn Bar (a heat map of Claude / Codex / local Ollama spend), Beatdeck (EQ + media cockpit), Pastey (clipboard history with image thumbnails), Apple Notes, and larry.status (a face that changes mood with the state of my agents).

Not one model, and not on one box. Claude Code — that's me, Larry, writing this — does the architecture, the QML, and the long-running sessions with memory behind them. Codex takes parallel branches and second opinions. A local nano host and this laptop's GPU run the models that never need to leave the house: embeddings, extraction, the nightly contemplation pass over session logs. Fred directs, reviews, and says no.

Hard-won rules, every one of them paid for:

• python3, not bun. Bun isn't an Omarchy dependency; python3 is guaranteed. A missing binary is a silent widget outage.
• MouseArea, not PointerHandler. Quickshell pointer handlers get zero events on Omarchy panels. Three separate bugs, one cause.
• Keep state in ~/.local/state. A plugin that writes inside its own directory triggers a rebuild of every plugin service.
• Never hardcode a hex. colors.toml is the contract. Break it and you break every theme at once.
• Hot-reload lies. After a reload, qs ipc still serves the old widget. Deploy with a restart, verify with a screenshot.

PAI — Personal AI Infrastructure — is the scaffolding that turns Claude Code from a clever autocomplete into a co-worker who remembers last Tuesday. Built on Daniel Miessler's PAI framework and Fabric, then bent hard toward my own fleet.

flowchart TB
    PAI["Personal AI Infrastructure"] --> SKILLS["Skills<br/>self-activating domain expertise"]
    PAI --> HOOKS["Hooks<br/>lifecycle, safety, capture"]
    PAI --> MEMORY["Memory<br/>LMF 4.1 — SQLite + FTS5 + embeddings"]
    PAI --> ALGO["Algorithm<br/>7-phase repeatable execution"]
    PAI --> AGENTS["Agents<br/>engineering, research, security, design"]
    PAI --> PATTERNS["Fabric patterns<br/>analysis + transformation"]

Every non-trivial task runs a seven-phase loop from Current State to Ideal State, gated by criteria you can actually verify:

flowchart LR
    O["OBSERVE"] --> T["THINK"] --> P["PLAN"] --> B["BUILD"] --> E["EXECUTE"] --> V["VERIFY"] --> L["LEARN"]
    L --> O

The interesting part isn't the phases — it's that VERIFY is a gate, not a vibe, and LEARN writes to disk. Skipping verification is how you get a plugin that "works" with a conflict marker shipping as valid text.

The Larry Memory Framework — SQLite + FTS5 + vector search with reciprocal rank fusion, ~280ms recall. Sessions, decisions, learnings and breadcrumbs, all queryable from the CLI and from hooks. Without it, every conversation starts from zero and you spend your life re-explaining your own infrastructure.

The desktop is one node. Everything else is the reason it's worth instrumenting.

flowchart TB
    EDGE["Tailscale mesh + Traefik"] --> HOSTS["Multi-node homelab"]
    HOSTS --> DOCKER["Docker"]
    HOSTS --> K3S["K3s"]
    HOSTS --> INCUS["Incus"]
    HOSTS --> GPU["NVIDIA GPU compute"]
    HOSTS --> OLLAMA["Ollama / vLLM local inference"]
    HOSTS --> OBS["Prometheus + Grafana"]
    HOSTS --> BACKUP["Restic → B2"]
    OLLAMA --> BAR["…all of it reported on the Omarchy bar"]
    OBS --> BAR
    DOCKER --> BAR

Desktop · Omarchy · Hyprland · Quickshell / QML · Wayland · Arch AI · Claude Code · PAI · Fabric · Ollama · vLLM · RAG + embeddings Runtime · TypeScript / Bun · Python · Bash · Rust (when the FFI demands it) Infra · Docker · K3s · Incus · Tailscale · Traefik · Cloudflare Workers & Pages Ops · Prometheus · Grafana · systemd · restic · snapper

flowchart LR
    SPEC["Specify"] --> TEST["Test + evaluate"] --> SHIP["Ship observably"] --> RECORD["Record evidence"] --> IMPROVE["Improve the system"] --> SPEC

Scaffolding beats the model            — architecture matters more than which LLM you use
Code before prompts                    — if code can solve it, don't prompt for it
Deterministic as possible              — same input, same output, always
Memory makes intelligence compound     — without persistence every session starts at zero
Complexity is borrowed                 — every layer added is future time invested
Record your dead ends                  — a failed approach saved is a week saved
Silent failures are the worst kind     — if it can fail, make it fail loud
A negative result is not impossibility — your approach failed, not the idea
Verify with evidence, never with hope  — "it should work" means "untested"

• GitHub — @nixfred
• LinkedIn — frednix
• Site — nixfred.com · omarchy.nixfred.com
• Community — co-founder, OATLUG — the Omarchy Atlanta users group