Vertreel
Watches a live stream, finds the moment worth keeping and renders it as a captioned vertical reel.
- 3
- platforms watched: Twitch, YouTube, Kick
- 90 s
- learning period before a stream can trigger a clip
- 12
- maximum clips per hour, per stream
- 4
- plans, from free to studio
01Summary
Vertreel watches live streams, detects highlight moments and renders captioned 9:16 reels, each with a title, description and tags ready to paste for every platform. In public mode it needs only a channel link. The creator does not have to log in.
It is a multi-tenant web app that also installs on a phone. Workspaces have their own streams, clips, visual style and plan, and one watched stream is recorded and analysed once however many workspaces follow it.
02Why I built it
A stream runs for hours and the part worth sharing lasts thirty seconds. Finding it by hand means watching everything twice. I wanted the detection to come from the audience itself: the chat and the audio already tell you when something happened.
03How it works
- Step 1
Watch
Follow the stream and its chat. Learn what normal looks like for this particular channel.
- Step 2
Detect
Three signals: chat speed against the stream's own baseline, many different chatters using clip, laugh or hype words, and sudden loudness.
- Step 3
Judge
Transcribe the best candidates and have a local model rank them. Audio-only moments must clear a higher bar.
- Step 4
Cut
Snap the start and end to sentence boundaries so a clip never opens mid-word.
- Step 5
Render
Crop to vertical following the action, add captions, facecam or B-roll layout, watermark.
- Step 6
Write
Draft the title, description and tags for each platform.
Fig. 2Videos readable out of 12, by where the request came from
Measured while building the learning loop. From a datacentre address YouTube asked every request to sign in. I did not try to evade it. Collection moved to my own machine and only the feature rows are synced to the server.
04Decisions
| Decision | Reason |
|---|---|
| Free by default | Copy is written by a local model first, then free tiers, then a plain template. A paid model is used only if switched on. |
| One chatter counts at most twice | A single excited viewer should not be able to trigger a clip alone. |
| Creative Commons B-roll only | The licence field is verified. A title that says no copyright is ignored. The source is credited on the video and in every post. |
| A small ranker, not fine-tuning | The learning loop is a logistic regression checked on held-out videos. Its labels are proxies for views, and the documentation says so. |
05What broke
- The ffmpeg build had no subtitle filters, so captions are drawn as image overlays instead.
- Kick keeps its chat protocol private, so Kick clips come from audio alone.
- A true live shadow test would have used up the stream slots, so it is deferred and documented as not built.
06What is next
Payments review and a first round of real users.