Skip to content
all workflows

case study · multi-agent system

AI Reel Pipeline case study

Producing short-form vertical video by hand is slow and repetitive. → LangGraph agents script, plan and render vertical reels end-to-end with Remotion.

private repo · case study only
TopicSupervisorScript +HookVisualprompt +AssetEditingagentsQCRemotionrenderoutput · vertical reelrendered reelplaceholder · 9:16
  1. 01 · The problem

    The problem

    Producing short-form vertical video by hand is slow and repetitive.

  2. 02 · My role

    My role

    Designed and built at Aevia.

  3. 03 · Architecture

    Architecture

    1. Topic

      one topic in, finished reel out

      → Supervisor

    2. Supervisor

      LangGraph · routes the workers

      → Script + Hook

    3. Script + Hook

      beat-by-beat script, scroll-stopping first beat

      → Visual prompt + Asset

    4. Visual prompt + Asset

      generates the images

      → Editing agents

    5. Editing agents

      parallel: camera, transitions, pacing, grade, text, voiceover, music, SFX, beat sync

      → QC

    6. QC

      validates everything; render runs only if it passes

      → Remotion render

    7. Remotion render

      video defined in React

      → Vertical reel

    8. Vertical reel

      1080×1920 MP4 · 30 fps · voiceover + animated captions

  4. 04 · Key decisions

    Key decisions

    • LangGraph: an explicit graph with a supervisor, parallel branches, merge steps and a conditional QC → render edge. State is shared and inspectable.

    • Remotion: the video is defined in React code, so motion, transitions and captions are fully programmatic and repeatable, with no manual editing.

  5. 05 · Challenges

    Challenges

    • Timing: fixed short beats clashed with the real voiceover length, so captions were unreadable. Beat duration is now driven by the voiceover length.

    • Parallel agents overloaded the LLM endpoint, so calls now retry with backoff.

  6. 06 · Result

    Result

    The full pipeline runs end-to-end on a server and renders a finished reel from a single topic. Private repo, case study only.

  7. 07 · Stack

    Stack

    • LangGraph
    • Multi-agent
    • LLM APIs
    • Remotion