TrueView

Banner
Main Screen
Ai Generated photo/audio results
Figma
Human photo/audio results

Inspiration

AI-generated photos and cloned voices get harder to spot every day. I wanted a detector that's always with you, your phone, and doesn't upload your private media to a server to do its job.

What it does

TrueView is an iPhone app that detects AI-generated photos and 4-second voice clips, entirely on-device. Snap or upload a photo, record a voice clip, and get a clear verdict, Bloom-Blessed (real) or Blight-Touched (AI-generated), with a confidence score. Your media never leaves your phone.

How I built it

Swift/SwiftUI app with AVFoundation for audio recording and PHPicker for photo selection. I converted Hugging Face transformer models to CoreML with coremltools, verifying that the on-device model matches the original PyTorch version. Custom preprocessing pipelines replicate each model's exact training preprocessing, resize method, normalization values, all of it.

Challenges we ran into

Model selection was brutal. My first model flagged everything as fake; the second couldn't spot AI images at all. I learned to verify empirically, testing in Google Colab, checking confidence scores, curating test sets and got guidance from the Hugging Face team right here at Knight Hacks. I also discovered that re-compressed images can flip a verdict, and that the iPhone's preprocessing must match the model's training pipeline exactly or accuracy silently breaks.

Accomplishments that we're proud of

Fully on-device inference, no server, no API calls, works in airplane mode. Photo AND voice detection in one app. A verdict UX (Bloom-Blessed / Blight-Touched) that makes deepfake detection feel approachable instead of intimidating.

What we learned

How to convert PyTorch models to CoreML and prove the conversion is faithful. That deepfake detection is as much about the data pipeline as the model itself. And to evaluate models empirically instead of trusting download counts.