Full-stack architecture

Now that you know why local Whisper matters, let us map the pieces you will wire together. The app stacks a React recorder in the browser, a FastAPI backend, an on-device Whisper model, and an optional LLM cleanup step.

Two reasons. First, users sometimes want the raw transcript without LLM cleanup, for example when they trust their own speech. Second, if the LLM fails or is slow, you can still show the Whisper output immediately and clean it asynchronously. Keeping the steps independent makes the pipeline resilient.