Inspiration
Our inspiration came directly from our teammate Camille, who works in wardrobe for major TV productions including Apple TV+'s For All Mankind. Through her experience on film and TV sets, she's navigated the real, recurring headache of brand clearance — logos and product placements that slip through editing and cause legal headaches or costly reshoots. She helped us understand that this isn't a hypothetical problem: studios need a reliable way to catch brand exposure before it becomes someone's liability. We wanted to build something that would fit naturally into the workflows real post-production teams already use.
What it does
Brand Unsafe automatically scans video footage for potentially problematic brand logos and product placements. It flags frames containing visible brand marks, giving productions a head start on clearance review before content goes to legal or distribution. The tool integrates directly with Frame.io so editors and supervisors can work within a platform they already know, reviewing flagged moments without breaking their existing workflow.
How we built it
We built Brand Unsafe as three connected layers: a front-end interface for viewing and managing flagged results, a Frame.io integration that hooks into the review and collaboration workflow productions already use, and a backend that handles video ingestion and runs our detection pipeline. (Add a sentence here about your specific model/stack — e.g. what video model or vision API you used for detection, and how your backend processes the frame data.)
Challenges we ran into
Getting the Frame.io integration talking cleanly to our frontend was our biggest technical hurdle — syncing state between the two platforms required more plumbing than we anticipated. Web deployment surfaced its own set of configuration headaches. On the ML side, coaxing reliable, structured output from our detection pipeline took significant prompt and post-processing iteration; getting clean JSON we could actually render in the UI was trickier than it looked.
Accomplishments that we're proud of
We built something that solves a real production problem, scoped and validated by someone who lives it every day. Getting the Frame.io integration working end-to-end — from footage upload through flagged frame review — felt like a genuine milestone. We're also proud that the core detection actually works: it surfaces real brand exposure rather than flooding reviewers with noise.
What we learned
We came away with a much deeper appreciation for video foundation models and what's now possible with vision-based detection at the frame level. We also learned how much of real product work lives in the integration layer — connecting APIs that weren't designed to talk to each other is often harder than the core ML problem. And Camille's domain knowledge reminded us how valuable it is to build with someone who knows the problem from the inside.
What's next for Brand Unsafe
The immediate priority is polish — tightening the UI, improving detection accuracy, and reducing false positives. Beyond that, we want to go deeper into editor workflows: tighter timeline integration, automated timestamped reports that could plug into clearance documentation, and potentially expanding detection beyond logos to products, signage, and other clearable IP.
https://github.com/markharmon868/TwelveLabsHackathon-BrandCompliance
Log in or sign up for Devpost to join the conversation.