docs/roadmap.md

# Roadmap

See [[final-build-plan]] for the detailed implementation plan.

## Phase 1: Refactor Without Behavior Change

- Split `app.py` responsibilities into focused modules:
  - `models.py`
  - `storage.py`
  - `transcription.py`
  - `llm.py`
  - `markdown.py`
- Keep the existing single-recording endpoint working.
- Keep Hugging Face deployment working.

## Phase 2: Meeting Session MVP

- Host creates meeting id and host key.
- Host receives participant join link and private host link.
- Participant joins with name label and recording mode.
- Participant submits audio segment.
- Host sees segment list.

## Phase 3: Room Recorder Support

- Add room label.
- Add covered participants field.
- Render Room Recorder transcripts separately.
- Warn if in-person meeting has no Room Recorder.
- Warn if multiple Room Recorders may duplicate speech.

## Phase 4: Transcript Correction and MoM

- Record or upload audio.
- Transcribe with Whisper `base`.
- Correct transcript using meeting context.
- Generate MoM.
- Save raw transcript, corrected transcript, and MoM as Markdown.
- Provide download buttons.

## Phase 5: Hugging Face Persistence

- Store meeting Markdown outputs outside ephemeral Space disk.
- Preferred option: Hugging Face Dataset repo.
- Alternative options:
  - GitHub repo
  - Supabase
  - S3/R2-compatible object storage

## Phase 6: Host Dashboard Polish

- Copy join link.
- Copy host link.
- Regenerate final MoM.
- Download Markdown bundle.
- Close meeting.

## Phase 7: Optional Speaker Intelligence

- Add deduplication for same-room duplicate recordings.
- Evaluate diarization only after labelled multi-device mode is stable.
- Consider `pyannote.audio` only if hardware, latency, and token requirements are acceptable.