While building InvoiceKit, I created deterministic verification paths alongside live Firebase regression checks for important invoice transitions. They answer different questions: deterministic paths make UI and navigation failures repeatable, while live checks confirm that the backend transition still works against the actual service. For client QA, name both layers in the scope. “Tested” is too vague unless everyone knows what was simulated and what integration ran live.

I herded geese as a kid.

There is a point in every developer's life where you either put a server in your apartment or go raise geese on a farm. I have done the geese already, so the server it is.

I set it up today on a machine I built in 2017 for a blockchain project. It cost me $2,700 back then. i7, 32GB of RAM, a 1070 Ti with 8GB. For nine years it has been alternating between sitting in a corner and doing real work, and the thing still holds up. It runs everything I am throwing at it now, which says more about how often we replace hardware than it does about the hardware.

The new project needs a TTS model. I added up what the infrastructure would cost through APIs and the arithmetic was not subtle. Hosting at home does not make development free, but it moves the number far enough that I stopped arguing with myself about it.

Meanwhile claude_mate is at 7 stars, 2 watchers and a fork. More interesting news soon.

github.com/vlgutv22/claude_mate