The lab

Five Proxmox nodes, two NAS boxes and a 100 GbE switch in a rack. Everything on it is defined in one repo — Packer builds the images, OpenTofu provisions the VMs, Ansible configures the hosts, and ArgoCD syncs the cluster — so the interesting question is never "what did I click" but "which commit did that". Most of the posts here start as something in this list breaking.

proxmox 9.2k3s 1.35.77-node cluster89 argocd apps100 GbE core287 GB VRAM283 TB raw

01 · Hardware

NodeSpecRole

pve1 – pve3i9-13900H · 96 GB · NVMe · 100 GbEk3s + compute

pve4i5-1240P · 16 GB · 2.5 GbEquorum

pve5EPYC 7313P · 256 GB · 2× RTX PRO 6000 + 3× RTX 4090 + RTX 3090GPU + desktops

pve-qnapVM inside the NAS hypervisorDR witness

bignasQNAP TVS-AIh1688ATX · QuTS hero · 12-wide raidz2 + NVMe poolbulk + S3

Synology DS1821+8-bay, separate power and switch portsecond copy

CRS520-4XS-16XQ16× 100 GbE QSFP28 + 4× 25 GbE SFP28, MTU 9000core switch

UDM SErouting, firewall, DNS, VLANsgateway

Raspberry Pi 5Zigbee + Z-Wave sticks, MQTT brokerradios

02 · What's running

ServiceWhat it doesState

Proxmox VEthe hypervisor cluster9.2.3

k3swhere apps landstable

ArgoCDapp-of-apps GitOps89 apps

CloudNativePGpostgres for 14 appsstable

Longhornreplicated volumesstable

Traefikingress and TLSstable

AuthentikSSO for everythingstable

Cloudflare Tunnelthe public hostnamesstable

LiteLLMone API for every modelstable

vLLMlocal model servingtuned by hand

SigNoztraces, metrics, logsstable

Prometheus + Grafanathe VM-side halfstable

Actions Runner Controllerself-hosted CIstable

Zotregistry + build cachestable

Immichphotosstable

Paperless-ngxscans and OCRstable

Plex + arr stackmediastable

TdarrGPU transcode queuefiddly

Home Assistantthe housestable

n8nworkflow automationqueue mode

CNPG backupsnightly to S3-on-HDDslow on purpose

Hand-updated when something changes. Not a status page — this blog is deliberately hosted somewhere else.

03 · Models

ModelWhat it's forSize

DeepSeek V4 Flash Visionreasoning, vision, 512k context285B-A13B MoE

Qwen3.8 27Beveryday chat and code, 256k context27B · AWQ INT4

QwenPaw Flash 9Btool-calling agents, always resident9B

Qwen3 Embeddingembeddings for RAG and search0.6B

Parakeet TDT v3speech to text0.6B

Kokorotext to speech82M

Served with vLLM behind llama-swap, which loads and evicts them across the cards on demand. A LiteLLM gateway puts these and the hosted frontier models behind one OpenAI-compatible API, so anything I build can move between them by changing a string.

04 · Coming soon

Still deciding what belongs here — most likely the decisions behind all of the above, including the ones I reversed.