5
Hosts
7
GPUs
158
Pods
102
Cores

Omarchy Homelab
Infrastructure

A five-node Kubernetes cluster with GPU compute, GitOps delivery, and centralized observability.

Scroll to explore

Mining Infrastructure Meets Cloud Native

The Challenge

Managing a 5-node Kubernetes cluster and 7 GPUs across mining, media, and AI workloads — with everything delivered from git and monitored centrally.

The Solution

An owned Omarchy fleet (Reverb-OS) plus ArgoCD GitOps for every cluster stack — one source of truth, no drift, no cloud spend.
102
CPU Cores
140GB
RAM
7
GPUs
9.5TB
Storage

Infrastructure Innovation

Git-Sourced Fleet

Host config lives in git (Reverb-OS: Omarchy + Home Manager). The live host is never the source of truth — changes ship from the repo, not from SSH.

Just-Driven Ops

One justfile per domain: just check, build, switch, deploy, health. The supported update path for the Omarchy fleet — and the fastest way back to a known-good state.

GitOps for Kubernetes

ArgoCD app-of-apps reconciles every cluster stack from git — media, Quill/MapleSpike, mining, Haven, and trading. Changes land by commit; rollbacks are git reverts.

Cluster Status

Kubernetes

5 nodes • 25 namespaces • ~160 pods
Every stack reconciled by ArgoCD — media, Quill, mining, Haven, trading.

AI Stack

Ollama on Zephyr (RTX 3090) and Nexus
Voice models + TTS on K3s; MCP server fleet wiring agents into every service.

Observability

VictoriaMetrics + Grafana + vmalert
GPU utilization, cluster health, and service SLO dashboards.

Operations & Automation

Automated Deployments

Every stack ships from git: commit, ArgoCD sync, rollout. Miner restarts are a git-native change, and workloads auto-restart on GPU availability.

Service Mesh

Calico CNI for pod networking, CoreDNS for service discovery, Longhorn for replicated storage. VictoriaMetrics + Grafana for monitoring and alerting.

Observability

Metrics with VictoriaMetrics (vmagent + vmsingle), dashboards in Grafana, alerting with vmalert. Logs via journald and pod logs.

Cluster Capabilities

Hardware Infrastructure

102 CPU cores, 140GB RAM, 7 GPUs (Zephyr: RTX 3090 + RTX 3060 Ti, Nexus: RTX 3060 Ti, Forge: 2x RTX 4060 + RX 5700 XT, Sentry: 1x AMD), 9.5TB storage. Mining runs as orchestrated K3s workloads across 5 GPUs; the cluster serves ~160 pods across 25 namespaces.

AI/ML Services

Local LLM inference with Ollama on Zephyr (RTX 3090) and Nexus; voice stack (voice-models + Chatterbox TTS) on K3s; MCP server fleet (arr, Jellyfin, qBittorrent, Quill) wiring agents into every service. No hyperscaler dependency.

Observability Stack

Metrics: VictoriaMetrics + Grafana
Logs: journald + pod logs
Alerting: vmalert rule evaluation
Dashboards: GPU utilization, cluster health, service SLOs

Infrastructure Timeline

The Journey

1
Before Sept 2025

Windows + Proxmox

Dual-boot setup with Proxmox servers for testing, Windows as daily driver

🎯
September 2025

Killed Windows

Full commitment to Linux. Started distro hopping journey.

3
Sept 2025 - Feb 2026

OS Evolution

Omarchy (Arch-based) → CachyOS (optimized Arch) → NixOS (declarative + reproducible)

4
March 2, 2026

NixOS Initial Commit

First NixOS configuration. Single host (zephyr) with basic desktop + gaming.

5
March 3, 2026

AI Gateway v1.0

OpenAI-compatible API, mining infrastructure, multi-GPU support

6
March 4, 2026

Gateway v2.0

Middleware architecture with circuit breaker, rate limiting, Redis caching

7
Mid-March 2026

Cluster Expansion

Added nexus, forge, sentry. Implemented NFS config sync, profile system, 50+ Justfile commands

8
March 18, 2026

K8s Phase 1-3: Foundation

Control plane, Flannel CNI, CoreDNS, stateful services (GlitchTip PostgreSQL)

9
March 19, 2026

K8s Phase 4-5: Services & GPU

Stateless services (GlitchTip web/worker, SearXNG, n8n), GPU workloads (llama.cpp)

10
March 20, 2026

K8s Phase 6-7: Monitoring Stack Complete

Prometheus + Grafana monitoring deployed for cluster-wide observability.

🚀
March 21, 2026

Production Cluster Live

60+ pods running across 4 hosts. AI inference, mining, and monitoring workloads in production.

12
April 1, 2026

Blog Content Collections

Migrated blog to Astro Content Collections. Type-safe content management with automated OG image generation.

13
April 16, 2026

Infrastructure Extraction

Extracted MCP Registry, Knowledge Fabric, and AI gateway into standalone declarative projects with reproducible config and CI pipelines.

14
April 17, 2026

Gateway V2.1

Connection pooling, response caching, K8s embed-server integration, and OCI container image builds for the AI inference gateway.

15
April 23, 2026

Code Quality Automation

Standardized pre-commit hooks (statix, deadnix, lint) across all infrastructure projects. Automated code quality enforcement.

16
May 9, 2026

Portfolio Design V2

16-theme retro gaming design system with Base24-compliant semantic tokens, theme-aware components across the entire portfolio.

17
May 18, 2026

Semantic Token Redesign

Theme architecture streamlined. Removed theme preview system, migrated to pure semantic tokens for maintainability.

🚀
May 19, 2026

Integration Roadmap Complete

Cross-codebase integration roadmap covering MCP ecosystem, knowledge pipeline, GPU federation, and unified auth across 12 repositories.

🐚
Aug 2026

ReverbOS / Omarchy Pivot

Full-stack migration off NixOS to an owned Omarchy-based OS across all hosts. Home-manager layer, declarative profiles, AI-first tooling. Sovereign infra future-proofing.

September 2026

Migration Complete

Every host migrated off NixOS to Arch-based Omarchy — K3s, AI inference, mining, and monitoring carried over without breakage.

🖥️
September 2026

Fleet Telemetry Desk

Infomarchy desk deployed to every host (fix merged upstream) — live AI sessions, mining, and system health across the fleet.

🔁
September 2026

GitOps Everywhere

Every cluster stack moved under ArgoCD app-of-apps management — media, Quill/MapleSpike, Haven, and trading deploy from git. No kubectl, no drift.

⛏️
September 21, 2026

Mining Joins the Cluster

All five GPU rigs migrated from native systemd services to K3s pods managed by ArgoCD (mining-helm). Zephyr joined as an agent to serve its two GPUs — rollback is one git revert.

🌐
September 2026

Self-Hosted, Full Stack

Community server (Haven) live at haven.reverb256.dev with a full Discord mirror; an autonomous agent fleet runs daily operations.

Code Explorer

Code Patterns

Overview.nix
# NixOS Configuration Pattern
{
  # Declarative system configuration
  boot.loader.systemd-boot.enable = true;

  # Network setup
  networking.networkmanager.enable = true;

  # User management
  users.users.jkro = {
    isNormalUser = true;
    extraGroups = [ "wheel" "networkmanager" ];
  };

  # System packages
  environment.systemPackages = with pkgs; [
    vim git curl wget
  ];
}