A voice deepfake fooled her grandfather. She built a startup that runs on the device.

When the call came, Tarini Padmanabhuni's grandfather was convinced he was talking to his brother. The voice on the line said he'd been kidnapped. The only way out: pay the ransom. The grandfather paid. Only later did they learn the brother was somewhere else entirely, unaware a scam was underway. The voice, it turned out, was a deepfake — an AI-generated clone. "What stuck with me wasn't the money," she says of the incident, roughly two years ago. "It was that my grandfather had no way to know."

That experience became DetectifAI, a San Francisco startup trying to make sure no one falls for the same trick.

The problem and the numbers

According to the FBI, Americans lost nearly $900 million to AI-based scams last year, up 24% from 2024. Victims 60 and older lost twice as much as those 50 to 59. The market is crowded — Reality Defender, Pindrop, Resemble AI, Microsoft Azure AI Content Safety, and Nuance (also Microsoft-owned) — but every existing product runs in the cloud on remote servers. That means phone makers can't embed them directly in the device, and the person under attack is left with almost no defense.

Technical approach: compact models on the device

Instead of shrinking large cloud models to fit a phone, DetectifAI designs compact, on-device models from the ground up — small enough to run inside a smartphone's operating system. The goal: an instant verdict on whether a voice is AI-generated during calls, voice messages, and other audio files, without the audio ever leaving the device. The company has not published public benchmarks or false-positive/negative rates, so production performance has not been independently measured.

Business model and first customers

The core is an SDK, a development kit that can be embedded in products and licensed through existing channels. The initial play: sell to phone makers, who get the detection as a built-in OS feature. Padmanabhuni's analogy is AT&T's role in launching the original iPhone — the first manufacturer to integrate DetectifAI wins a competitive edge. A second revenue stream is planned: licensing to enterprises and fraud-prevention firms. For now, she says, there's early revenue: more than 100 thousand calls a month to financial institutions in India, where AI voice agents handle debt collection and loan-document tracking, with deepfake detection and speaker verification on every call. Customer names are withheld under NDAs.

Founder background, funding, and stage

Padmanabhuni says she started in machine learning at 12, studied cyber-physical systems — tech that links software with physical machines — at the Manipal Institute of Technology in India, and was the youngest team lead in what she describes as India's first autonomous racing-car division in the international Formula Student competition. DetectifAI raised a small seed round from Josh Constine (former TechCrunch editor) and Manohar Kamath (partner at KM Growth). The company was selected by TechCrunch editors to compete in the Startup Battlefield at TechCrunch Disrupt, running October 13–15 at San Francisco's Moscone Center.