Skip to content
AIJul 17, 20262 min read

Ultimate Reel Maker — why I self-host the video pipeline

Topic to vertical MP4 on my own metal: script, TTS, captions, B-roll, and Remotion — without renting a SaaS editor.

VideoRemotionDockerBuild in public

The question

If I want a short-form video from a topic in under an hour, do I need another subscription — or can I assemble the pieces I already trust locally?

What I was optimizing for

Most "AI video" products hide the pipeline. You get a download button and no idea what failed when the voice sounds wrong or the captions drift.

Ultimate Reel Maker (Reelsmaker) inverts that:

  • Script from Ollama — I can swap models or prompts without renegotiating a vendor UI.
  • TTS via Kokoro — local, repeatable, no per-character billing surprises.
  • Captions from Whisper.cpp — word-level timing I can inspect.
  • B-roll from Pexels, uploads, or generated stills — not locked to one stock library.
  • Composition in Remotion at 1080×1920 — the layout is code, so "fix the title safe zone" is a PR, not a support ticket.

The bet is not "fastest demo." It is a pipeline I can debug at 11pm when one stage misbehaves.

How the pieces fit

The happy path is boring on purpose:

  1. Pick a topic and template.
  2. Generate narration and split into beats.
  3. Render audio, derive captions, fetch or generate visuals.
  4. Hand off to Remotion for the final vertical MP4.
  5. Drop the file into portfolio storage.

Docker keeps the services honest: web UI, API, Redis queue, worker, Ollama, and Remotion each have one job. When render fails, I know which container to read logs from.

What I refuse to outsource yet

  • Caption timing — if the words don't match the audio, the video is unusable; that check stays mine.
  • Final composition — Remotion code is the source of truth for brand-safe layout.
  • Human review — automation stops before "post to TikTok"; I still watch the cut once.

Mistakes I'm still fixing

  • Underestimating cold-start time when Ollama and Remotion spin up together.
  • Letting B-roll selection be too random — template beats need tighter constraints.
  • Treating "it rendered" as "it's good" without a one-minute watch pass.

Takeaway

Self-hosting video is not nostalgia. It is ownership of the failure modes. If you care about repeatability more than magic, assemble the pipeline yourself and diagram it so the next you knows where to look.

Architecture diagrams: Reelsmaker — system flows