Inspiration

Editing a video is only half the job. Right before you share it, you spot a face or an email address in the frame and have to start again. We wanted an editor that does the edit and the safety check in one place.

What it does

You upload clips and type what you want, like "a 30-second highlight, best moments first". Gemini turns that into an edit plan and FFmpeg cuts it locally. You can control it hands-free with gestures [check], and one button covers faces, emails and phone numbers in the finished video. The detection runs on your laptop, so nothing is sent to the cloud for it.

How we built it

A Node server with a browser editor. Gemini plans the cuts, a deterministic FFmpeg renderer does them, so the AI decides but can't break the video. For privacy we use a local face detector and OCR, then re-check the finished video as a safety gate.

Challenges we ran into

Testing on real footage showed our first version leaking text and a face for a split second when a window moved and a person walked out of frame. We fixed it by tracking things through gaps and adding a check that rejects the output if a face is still showing. Our first blur also looked fine in the numbers but too weak to our own eyes, so we switched to a solid fill. Phone video (rotation, HDR, odd frame rates) and slow text scanning on busy screens were the other headaches.

Accomplishments that we're proud of

  • It fails closed: if anything goes wrong, it deletes the output instead of handing over a video marked as redacted, and the download stays locked until you've watched the result.
  • We tested on real phone footage (side profiles, dim light, two people far away) and kept fixing it until the leaks we found were gone.
  • Everything privacy-related runs on the laptop. ## What we learned A detector checking its own work isn't real verification, so we put a human in the loop and say so in the interface.

What's next for Multi-purpose video editing

What's next for Cut the Brief

It can miss small, blurry or moving text and tiny faces, so always watch the result before sharing. Frames sent to Gemini for planning aren't redacted first yet. Next up: that fix, choosing which faces to cover, and a faster text scan.

Built With

Share this project:

Updates

Submission history