homelabs fail in a predictable way. you buy interesting hardware, stand up interesting software, and six months later you have a museum of almost-working services that only you can restart.

the antidote is to treat the lab like a product — a small one, with a clear purpose and an operating model that survives your own forgetfulness.

for me that means intentional topology: compute roles instead of random boxes, network segmentation instead of a flat trust-everything lan, storage that is planned for failure, and kubernetes that is reconciled from git rather than clicked into existence.

gitops in a personal lab is not theater. it is how you remember what you intended. flux (or any reconciler) turns “i think i configured that” into “the repo says this is the desired state.” that habit is exactly what i want muscle memory for when the stakes are higher at work.

local ai fits the same frame. ollama, routing layers, and agents are interesting — but only if they sit on a path you can explain: where inference runs, how models are pulled, what is exposed, and what stays private. curiosity without boundaries becomes another undocumented service.

none of this is an employer system. that distinction matters. the lab is a sandbox for learning and validation, not a shadow of production. the value is the practice: documenting purpose, automating the boring path, and refusing to let novelty outrun operability.

curiosity becomes craft when the lab has customers — even if the only customer is future you — and when the product is reliable enough that exploring the next idea does not require rebuilding the last one from memory.