Operator Retrospective · ~6 Months
From a handful of boxes to a self-hosted, self-monitoring AI homelab spanning multiple countries — running uncensored LLMs behind a Telegram agent, on its own private mesh network, with personal media streaming and offsite backups, and zero dependence on anyone else's cloud.
A multi-tier LLM stack on the GPU server — a Telegram agent with intelligent model routing, an uncensored deep tier, a sentiment sidecar, web search, and image generation. Private domains stay TUI-only by policy. Benchmarked and hardened through dozens of iterations.
The headline win: cut the cord from hosted Tailscale and stood up a self-hosted Headscale control plane — 7 nodes, server IPs preserved so nothing broke. Fixed the WAN underneath, incl. an uplink throttled 0.4 → 57 Mbps by our own shaper.
Jellyfin with GPU transcoding and VPN-isolated, killswitched torrents — evolved from tailnet-only into a public, WireGuard-tunneled stream you can watch from anywhere, at full uplink speed.
A 23-probe healthcheck plane with an out-of-band watcher, scripted fleet audits, an 8-agent security sweep (verdict: not compromised), everything critical in version control, and daily encrypted offsite backups.
The Punchline
It's not a pile of servers anymore — it's an operator-controlled, uncensored, resilient AI platform. It streams your media, answers on Telegram, guards your data, and depends on no external service to keep running. You own the whole stack, top to bottom.