Dovetail?

Build a focused replacement for the core Dovetail user-research workflow: a researcher imports interview recordings and transcripts (from Zoom exports or direct upload) into a searchable repository; the system transcribes audio with speaker labels, redacts PII, auto-tags themes, and lets teammates ask questions in natural language and receive cited answers drawn from actual transcript passages. AI then generates a stakeholder-ready summary or one-pager deliverable (Voice-of-Customer style) with quotes and source references, and optionally pushes a digest to Slack/email on a schedule.

User-facing requirements: (1) upload audio/video or paste transcripts; (2) live or file-based transcription with speaker diarization; (3) transcript viewer that highlights the audio position and lets users add notes/tags; (4) project-level search plus global natural-language Q&A returning answers with quoted evidence and clickable source snippets; (5) one-click generation of a findings summary/prd-style brief containing citations; (6) scheduled Slack or email digests of new synthesized insights; (7) role-based sharing (viewer vs editor).

Technical requirements: web app with a React/Next.js front end and a Node or Python API; Postgres with pgvector for embeddings and metadata; object storage (S3 or compatible) for media; speech-to-text API or Whisper-family model for transcription; an LLM API for tagging, summarization, and retrieval-augmented answer generation; queued background jobs (e.g., Bull or Celery) for transcription, agent schedules, and digest delivery; OAuth hooks for Slack and Zoom.

Implementation approach: use a RAG pipeline where transcripts are chunked, embedded, and stored in pgvector; answers are generated by retrieving top-k chunks and requiring the LLM to ground every claim in citations; use regex/sensitive-data detectors for initial PII redaction plus an LLM pass for transcript context; store all source-to-output mapping in a schema that makes auditable citations trivial. Seed the app with 2–3 sample interview datasets so the demo is immediately testable.

Do NOT copy Dovetail's actual brand, logo, visual design, UI layout, or name. Build an original UI and brand for this tool. Scope out: mobile apps, 75-language translation (support English plus 5 selectable languages max), enterprise SSO/compliance certifications, a marketplace, and beyond-text media analysis (focus on transcript-first functionality). Also scope out any system that claims to detect emotion; only silence/gap detection is in scope.

Deliver a README covering setup steps, environment variables, how to seed demo data, how the citation mechanism works, and important limitations: transcription/diarization accuracy varies with audio quality, LLM answers may mis-cite if chunks overlap, redaction is not guaranteed perfect, and embedding/LLM costs grow with repository size.

Out of scope for v1: enterprise-scale infrastructure, broad integrations, and feature parity with the reference product.