Can AI replace Fathom?

Build me a personal call-notes pipeline to replace Fathom. Requirements:

- A watched folder: I drop call recordings (m4a/mp3/mp4) into ~/Calls/inbox/,
  chokidar picks them up, a Node pipeline processes each one.
- Transcribe with whisperX for speaker-labeled, timestamped transcripts, or a
  hosted Whisper API if a key is present in .env.
- Summarize with an LLM (key in .env) using templates in ./templates/*.md. The
  default produces a 5-bullet summary, decisions, action items with owners, and
  open questions. Pick the template from the filename prefix (sales-, 1on1-).
- Clips: any transcript line can be cut into a share-able moment, ffmpeg trims
  15 seconds either side of the timestamp into ~/Calls/clips/.
- Output per call: ~/Calls/notes/YYYY-MM-DD-<name>.md, summary on top, full
  transcript below a divider.
- A small local web page (Express, localhost) listing calls with keyword search
  over transcripts.
- No accounts, no telemetry; audio never leaves my machine unless I opt into
  the hosted API.
- Out of scope: a bot that joins meetings, calendar auto-capture, CRM sync,
  team libraries. I record with my meeting app or OS and drop the file in.
- README: dependencies, keys, and how to capture system audio on macOS.