← all posts site ↗

building an ai console then finding lemonade

wanted one gpu accelerated app on windows to control every ai model on the machine. download tts, stt, and llm models, start and stop their endpoints, change ports, per service autostart toggle, logs, the whole thing. started building it in rust and immediately hit the classic trap: the start button worked before setup ran, so pressing it produced a missing module error instead of a server.

fixed it by making start do its own setup. it checks the venv, creates it, installs the packages, smoke tests the server, and only writes a ready marker after a clean install, so a half finished install can never look like a good one. reproduced the exact failure, wiped the venv, pressed start, and watched the card come up green thirty nine seconds later.

then came the better question: does something already do this? one app comes close, a tauri desktop app built by amd engineers that runs llm, stt, tts, and image from one install, with a model manager and one openai compatible endpoint. it maps perfectly onto the radeon gpu. but it is one server on one port, no per service ports, and stt only supports whisper builds, no parakeet. a global press and hold hotkey is a sixty line autohotkey script away because the api is open.

the takeaway is knowing when to stop building. the rust app works, the failure path is fixed, but the existing app is the better fit on this hardware, and the missing pieces are small clients, not a rewrite.