Project: Audio Transcription Pipeline

<!-- Hydrate with: "read plan.md and create tasks with dependencies" --> <!-- Last updated: 2026-03-04 | Session: zoom-pipeline-v2 -->

Status Legend

• pending — not started
• in_progress — currently active
• completed — done and verified
• blocked — waiting on dependency

---

Tasks

T1 · Set up S3 bucket and IAM roles

• status: completed
• blocks: T2, T3
• description: Create S3 bucket for raw Zoom recordings, output transcripts, and configure IAM roles for read/write access.
• notes: Bucket name: zoom-transcripts-prod. IAM role ARN saved to .env.example.

---

T2 · Build Zoom webhook receiver

• status: completed
• blocks: T5
• blockedBy: T1
• description: FastAPI endpoint that receives Zoom recording-complete webhooks, validates signatures, and enqueues the download job.
• notes: Deployed to Lambda. Webhook secret stored in AWS Secrets Manager.

---

T3 · Implement S3 download worker

• status: completed
• blocks: T5
• blockedBy: T1
• description: Worker that pulls raw .m4a files from Zoom cloud storage and stages them in S3 under raw/.
• notes: Handles Zoom OAuth token refresh automatically.

---

T4 · Integrate Whisper transcription

• status: in_progress
• blocks: T6, T7
• blockedBy: none
• description: Wrap OpenAI Whisper API (or local model) to transcribe audio files. Output raw transcript JSON with word-level timestamps.
• notes: Currently testing whisper-large-v3. Cost ~$0.006/min. Need to decide on local vs. API.

---

T5 · Wire webhook → download → transcription pipeline

• status: pending
• blocks: T8
• blockedBy: T2, T3, T4
• description: Connect the three components end-to-end. Webhook triggers download worker; on completion, enqueues transcription job.
• notes: Use SQS queues between stages for resilience.

---

T6 · Build diarization layer

• status: pending
• blocks: T7
• blockedBy: T4
• description: Add speaker diarization using pyannote.audio. Merge diarization output with Whisper timestamps to produce speaker-labeled transcript.
• notes: Requires HuggingFace token. pyannote model: pyannote/speaker-diarization-3.1.

---

T7 · LLM post-processing (cleanup + summary)

• status: pending
• blocks: T8
• blockedBy: T4, T6
• description: Pass labeled transcript through Claude to: (1) clean filler words, (2) generate structured meeting summary with action items.
• notes: Use claude-sonnet-4-6. Prompt template in prompts/summary.md.

---

T8 · Output delivery (email + Notion)

• status: pending
• blocks: none
• blockedBy: T5, T7
• description: Deliver final transcript and summary via: (1) email to meeting host, (2) Notion page in shared workspace.
• notes: Notion API token in Secrets Manager. Email via SES.

---

Dependency Graph

T1 (completed)
├── T2 (completed) ──────────────────────┐
└── T3 (completed) ──────────────────────┤
                                         ▼
T4 (in_progress) ──────────────────── T5 (pending) ──┐
├── T6 (pending)                                      │
│   └── T7 (pending) ──────────────────────────────── T8 (pending)
└──────────────────┘

Session Notes

• 2026-03-01: T1, T2, T3 completed. T4 started — testing Whisper models.
• 2026-03-04: T4 still in progress. Blocked on local vs. API cost decision. T6, T7, T8 not yet started.

Hydration Instructions for Claude

When asked to hydrate this plan:

1. Create a Task for each Tx entry
2. Set status from the status field
3. Wire addBlockedBy from each task's blockedBy list
4. Skip creating tasks with status completed — just note them as context
5. Begin work on tasks whose blockedBy tasks are all completed or absent