I built an auto-clipping tool from two pieces: xAI’s Grok speech-to-text, which returns a transcript with word-level timestamps, and Jev, TypeSafe’s classifier, which scores every candidate clip.
Jev splits the transcript into segments, asks whether each one is surprising or insightful, and ranks them. On one of my videos it returned 11 ranked clips in about two seconds.
Summary
- Jev outputs numbers and rankings, but its input can be anything you can express as text and questions.
- A larger model like Sol or Astra might pick a slightly better clip, but not in two seconds.
- Because it runs in real time, you can rephrase questions and re-evaluate the same input again and again.
- TypeSafe’s Doom demo shows Jev choosing actions like “hold fire” from shifting probabilities and chaining decisions into a decision tree.
How to apply it
- Clip long videos for TikTok and Shorts: transcribe with word timestamps, have Jev rank the segments, then cut the top results.
- Screen signups or incoming API requests for fraud with a detached Jev check that runs alongside each request.
- Score outgoing email on a sending platform for scam risk before it goes out.
- Build game or agent logic as chains of small decisions instead of one large prompt.
- Explore Jev or read the documentation, and leave your build ideas on the video.
Get the next test the day it’s out.
