Can AI replace tl;dv?

Build me a personal meeting recorder with clips to replace tl;dv. Requirements:

- A macOS menu bar app (or a rec start / rec stop CLI, pick the simpler to ship)
  that records system audio + mic locally with ffmpeg. No bot joins the call.
- On stop: transcribe with whisper.cpp with timestamps, or the Whisper API if a key
  is in .env.
- Send the transcript to an LLM (key in .env) for a 5-bullet summary and action
  items; save everything to ~/Meetings/YYYY-MM-DD-<title>/ as notes.md, the
  transcript, and the audio file.
- While recording, a global hotkey drops a highlight marker; afterwards each marker
  becomes a clip, ffmpeg cutting 30 seconds either side into clips/.
- A local library page (Express on localhost:4700): list meetings, full-text search
  across transcripts with SQLite FTS5, play recordings and clips inline.
- No accounts, no telemetry; nothing leaves my machine except optional API calls.
- Out of scope: a bot that auto-joins Zoom/Meet, CRM sync, and a shared team
  library. This records what I can hear locally.
- README: mic + screen-recording permissions, whisper model download, and a
  reminder that recording other people still requires their consent.