LazyCut

Turning a talking-head recording into a tight vertical short takes time and skill: trimming dead air, cutting filler words and fluffed takes, framing for 9:16 and adding captions. The alternatives were editing it yourself or paying an editor. The brief: let AI do the whole first pass from raw footage, while keeping every AI decision reviewable and adjustable before render.

At the centre is the EDL (Edit Decision List), a JSON document that is the single source of truth: the AI proposes edits to it, the user adjusts them, and the renderer reads the same file.

1. 1. Ports and adapters

   All editing logic lives in pure shared packages with no infrastructure attached, and every outside provider sits behind a swappable adapter. The web app is Next.js on Vercel; processing runs on a DigitalOcean worker, with Postgres and Drizzle, Spaces for media, and a Redis and BullMQ job queue.

2. 2. The auto-edit pipeline

   The worker extracts audio with FFmpeg, transcribes it word by word with timings, finds silences and filler words, and has an LLM — Claude, GPT or Gemini — plan the cuts, snapped to word boundaries so names are never clipped mid-word. It then places emoji on keywords, sound effects at cut transitions and one or two memes, deterministically and removably.

3. 3. A review-first studio

   A CapCut-style browser editor opens on the full take: the cuts the AI proposes are highlighted on the timeline to keep or cut take by take, and the subtitle list is split into sections labelled good take, bad take, dead air or filler, each one a click from its spot in the preview. A multi-track timeline carries B-roll you can position and scale.

4. 4. Render what you saw

   Remotion renders 1080×1920 video on the worker from the same components as the editor preview, tiling cuts frame-accurately so no one-frame blip appears between them, with IBM Plex Sans Thai loaded into the export.

5. 5. Nothing unproven switched on

   The auto-reframe stage is wired into the pipeline but stays off until a face tracker passes a real render test, and the credits-and-billing engine is built but not yet switched on.

The web studio and the processing worker are live in production.

• A first cut from raw footage

  Upload a talking-head clip, podcast or screencast, and the AI transcribes it, proposes cuts and adds captions in one pass.

• Every AI change stays editable

  Cuts, captions, emoji, sound effects and memes can all be adjusted or removed in the studio, and nothing is cut until the user approves it.

• Word-by-word captions

  Captions follow speech word by word in five styles, with adjustable font and position, and Thai renders correctly in the export.

• Ready for vertical feeds

  Exports a 1080×1920 vertical video for TikTok, Reels and Shorts.