Redactly

Inspiration

Every day, millions upload TikToks without realizing they might be exposing personal details — a license plate, a child’s face, an email on-screen, or even a phone number spoken out loud. Once it’s online, it’s too late. We wanted a privacy-first layer that empowers creators to stay safe without slowing down their creativity.

What it does

Redactly automatically scans your video before upload, looking at:

Visuals: detects faces, vehicles, license plates etc.
Speech & Captions: transcribes audio to catch phone numbers, addresses, names, etc using NER. It then surfaces all potential privacy leaks in an interactive UI where you decide: blur faces, pixelate plates, or bleep/mute sensitive words. With one click, it outputs a redacted video ready for TikTok.

How we built it

Computer Vision: YOLOS for people/vehicles, plus regex/NER for PII.
Speech: Whisper ASR pipeline to catch private info in dialogue.
Fusion Layer: merges detections across modalities, removes duplicates, assigns severity.
Frontend: Lynx UI that previews flagged regions and lets creators toggle actions.
Backend: FastAPI pipeline that scans, fuses results, and re-encodes the redacted video using OpenCV + ffmpeg.

Challenges we ran into

Performance: running multiple models per frame was slow, so we introduced frame sampling.

Synchronization: aligning millisecond timestamps across audio and video tracks for smooth redaction. Audio/Visual artifacts that were removed using FFMPEG

Accomplishments that we're proud of

Built a working end-to-end system: upload video → scan → interactive review → download safe version.
Designed a clean UI that empowers creators to stay in control instead of enforcing automatic censorship.