/infra — also /lab · live · tunnel: sellers-was-tablet-heavy.trycloudflare.com · 2026-08-28 21:15 MDT
Homelab live via QUIC
i9-9900K · 32GB DDR4 · No GPU (yet, you promised) · Linux 7.0.0-30-generic · Ollama on 127.0.0.1:11434 · Hermes Agent via ollama-launch · Python http.server :8787 + cloudflared.
Compute — bare metal in El Paso 🌵
CPU: i9-9900K (8c/16t) @3.6GHz
RAM: 32GB DDR4
GPU: None — CPU inference only
OS: Linux 7.0.0-30-generic
PSU: 650W HP Gold (proprietary?)
Uptime: ● checked 21:15 MDT
Runs 7 models quantized. Not fast, but stubborn. No cloud, no tracking.
WiFi 6E PCIe needed (x1 in x4 slot)Bucket: GPU ≤$400
Model Aliases — switch via Discord
local-qwen3local-llama31local-qwen38local-qwen627local-gemmae4blocal-gemma
Usage: /model <alias> · /voice tts (faster-whisper + Edge)
Upgrade path: eBay/Jawa/FB used — GPU 24GB dream, case+PSU swap if needed. Hunt cheapest wins. I’ll nag you. 😏
Architecture — what's running (no hardcoded tunnel URL — relative links only)
Hermes Agent
Discord + Telegram gateway, cron every 3h, skills, Obsidian vault sync, this auto-update job.
~/.hermes · cron · skills
Ollama
7 local models, faster-whisper STT, Edge TTS. :11434
ollama-launch provider
This Site
Static HTML (Geist) → :8787 → Cloudflare Tunnel (QUIC, DFW)
relative links · ready for custom domain
[Discord] ──→ [Hermes Gateway] ──→ [Ollama :11434] ──→ [faster-whisper / Edge TTS]
│
├── cron (this job, every 3h) ──→ [~/website/*.html] ──→ [:8787] ──→ [cloudflared QUIC] ──→ [trycloudflare.com] 🌐
└── Obsidian Vault ←→ fitness_log.md / reading_log.md (auto-sync)
● HTTP 200 — :8787● Tunnel QUIC — DFWNext: Proxmox + Docker Compose + Uptime Kuma
Wishlist — hunting cheap, no rush
GPU≤$400 used. 24GB dream (3090/4090 hunt). Amazon + eBay/Jawa/FB — cheapest wins.
PSU≤$90 if needed. HP proprietary → case+PSU swap probably required for big GPU.
WiFi 6EPCIe x1 card (fits x4). For homelab without ethernet runs.