A scam and deepfake detection engine built for the UC Berkeley AI Hackathon, and the closest a weekend project of mine has come to becoming a company.

Vify analyzes text, image, audio, and video together: an ensemble of a CNN for imagery, a Bi-LSTM for text, and audio networks, with the Hume API layered in for emotion analysis, reaching 79% detection accuracy across modalities.

We finished as finalists, were named one of the ten most investable ideas at the hackathon, and went into pre-seed term-sheet discussions with SkyDeck, Berkeley’s accelerator.

The uncomfortable lesson: detection is an arms race, and the defender’s model ages like fruit. It sharpened how I think about every AI product claim I read now, including my own.