Drop a recorder's files into an inbox and the pipeline takes over: wait for the file to stop growing, extract and flatten the mid channel from mid/side, diarize, transcribe with word-level timestamps, merge speakers, and archive the result with acoustic bookmarks intact.

Enrichment and chat run against a local model over Metal. The design rule is absolute and shapes everything else — no audio, no transcript, and no derived text ever leaves the machine, which rules out every hosted transcription service and is the reason the pipeline exists at all.