Coda

Coda: Meet, explore in 3D, and leave with everything auto-created.

Comment

Coda — Meetings That Generate Themselves

Inspiration

Every meeting ends the same way: someone says “I’ll send notes” — and they rarely do.
We wanted the outputs of a conversation — recap, deck, quiz, even a podcast — to exist automatically the moment the call ends.

At the same time, we questioned the format of meetings themselves. Why are we still stuck in linear slides and screen sharing?
Coda reimagines meetings as interactive 3D spaces, inspired by world models — where information is explored, not presented.

What It Does

Coda is a meeting platform that captures, understands, and generates everything from your conversation.

During the meeting:

  • Live transcription captures everything said
  • Screen capture adds visual context
  • Participants collaborate inside a shared interactive 3D space

The moment the call ends:

  • A recap page is instantly generated
  • Nine downloadable artifacts are created in parallel:
    • Report
    • Slide deck
    • Quiz
    • Flashcards
    • Mind map
    • Infographic
    • Video

How We Built It

Meeting Layer

  • Next.js + LiveKit for video infrastructure
  • Custom pre-join and room shell
  • Shared 3D worlds built with Three.js
  • Synced cursors and camera via a postMessage harness inside iframes

Capture Pipeline

  • Browser-based speech transcription
  • Periodic screen-frame uploads
  • Continuous streaming into a content backend during the meeting

Generation Backend

  • FastAPI service wrapping notebooklm-py
  • On meeting end:
    • Creates a notebook
    • Ingests transcript + visual context
    • Fans out parallel generation across nine artifact pipelines

Recap UI

  • Polling-based status page
  • One-click downloads for all artifacts
  • Landing page listing all past sessions

Challenges We Ran Into

  • Coordinating three separate apps (meeting, backend, recap UI) without a database

    • Solved using a lightweight JSON run log

  • NotebookLM client reliability

    • Browser-driven (Playwright), fragile authentication
    • Unpredictable generation timing

  • Embedding 3D scenes inside a live video call

    • Required heavy CSS and layout adjustments
    • Maintaining controls and synced cursors was non-trivial

  • Design transformation

    • Moving from default LiveKit UI to a distinctive Coda aesthetic without rewriting core components

Accomplishments We’re Proud Of

  • True end-to-end flow:

    • Leave a meeting → land on a styled recap page → download real artifacts
    • Formats include .pdf, .mp3, .mp4, .md, .csv, .json

  • Parallel generation system:

    • Faster artifacts are ready instantly
    • Heavier ones (audio/video) complete while users explore

  • Real-time shared 3D environments:

    • Synced cursors
    • Host-controlled camera
    • Embedded content panels inside scenes

  • Strong product identity:

    • Fraunces + Manrope + JetBrains Mono typography
    • Ember accent palette
    • Scanline grain + animated glyph branding

What We Learned

  • Parallel generation is a UX superpower
    Users engage immediately while heavier outputs catch up

  • Persistent visual cues build trust
    A subtle “recap arming” indicator is more effective than explanation

  • 3D spatial context changes meetings fundamentally
    Even a single shared room alters how people present and interact

  • Bold design decisions win
    Clear, opinionated aesthetics outperform safe, neutral design

What’s Next for Coda

  • Persistence & accounts

    • Postgres-backed workspaces
    • Recaps tied to teams, not JSON logs

  • Live mid-meeting generation

    • Reports and quizzes form in real time
    • Outputs appear inside the 3D environment

  • Scene editor

    • Teams create custom environments
    • Lecture halls, design studios, war rooms

  • Personalized artifacts

    • Each participant gets a tailored recap
    • Based on their contributions and questions

  • Export bundle

    • One click → full zip + static recap site

  • World model integration

    • Deeper semantic understanding
    • More intelligent, structured outputs

Built With

Share this project:

Updates