Bootstrapping Flux onto a live cluster is different from starting fresh. Reconciliation will fight workloads you deployed manually unless you annotate them first.

16GB RAM is enough for 8B models. Phi-3 Mini runs fine. Llama 3 70B needs a GPU. The quantized 4-bit version is barely usable at homelab speed.

Both have APIs. The sync logic was straightforward; the OAuth refresh for Toggl took most of the time. Claude wrote 90% of it. I spent my time on the refresh handling and the rate-limit edge cases.

Google Takeout exports cleanly. The problem is metadata: Takeout separates JSON sidecar files from images, and Immich's importer doesn't always reconcile the timestamps correctly. exiftool fixes it in bulk.

GridPane sends emails for maintenance and outages. I don't check email fast enough. Simple webhook from a scheduled script polling the status page; now it pings Slack immediately. Took an afternoon.