k3s · two nodes · qwen2.5 7b local
Conversation practice against a local 7B model
Writing correction with typed error kinds
Decks generated from the corpus, with audio
Carioca slang database, fed by coach output
Samba, MPB, choro
Podcasts and audiobooks
Films, with transcript breakdown tooling
Node and pod metrics, k3s-adjusted scrape config
CPU inference, 20Gi limit, measured 4.6 tok/s
The exercise said not to write the Ingress myself. Make the Helm chart do it. I spent the first half convinced I was just typing commands for no reason, and the second half fighting the space bar.
My bookmark manager only existed behind a port-forward. Getting it onto the public internet took a Cloudflare Tunnel, a Service I never made, and a certificate lesson I did not see coming.
The first course worked. The second one showed me which parts of my design were real and which parts were just how FSI happened to be laid out.
A free FSI course, an OCR text layer that is mostly right, and a five stage pipeline you can run twice without breaking anything.
One command, two numbers, and every one of them wrong at some point while I tested Navidrome and Open WebUI.
Two settings that had been quietly ignored for 37 days, a root-owned volume on the Dell, and the restart test that almost scared me.