DetectifAI Pitches On-Device Deepfake Voice Detection at TechCrunch Disrupt

Tarini Padmanabhuni launched DetectifAI, a San Francisco company that builds small AI models capable of spotting cloned voices on a phone itself, after a deepfake call targeting her grandfather. The scammer used a synthetic version of his brother's voice, and the con worked. DetectifAI is now one of the startups competing in the Startup Battlefield at TechCrunch Disrupt.

The mechanics of the fraud are familiar to anyone tracking voice cloning. A short sample of a relative's speech, often pulled from social media videos or voicemails, is enough for commercial tools to produce a convincing replica. Scammers then call with an urgent story about an accident, an arrest or a lost wallet. Padmanabhuni's grandfather heard a voice he trusted and acted before checking with anyone else in the family.

DetectifAI's answer is to keep the analysis on the device. Instead of shipping audio to a cloud server, the company's models are compressed enough to run on ordinary smartphones and flag synthetic speech as a call happens. That design cuts latency, keeps conversations off remote servers, and works when a phone has a weak signal. Cloud-based detectors can take seconds to return a verdict, which is too slow when someone is being pressured to send money.

Running detection locally is a harder engineering problem than it sounds. Voice models have to stay accurate while shrinking to fit within the memory and battery limits of a handset, and they must handle noisy calls, accented speech and the compression artifacts that phone networks introduce. Competitors including Pindrop, Reality Defender and several university labs have built strong cloud detectors, but on-device inference remains a smaller field. DetectifAI is betting that proximity to the call is the advantage that matters.

The Startup Battlefield slot gives the company a public test in front of investors and a live audience, and voice fraud is a receptive market. Regulators in the United States and Europe have started treating cloned audio as a consumer-protection issue, and banks have begun warning customers about family-emergency scams specifically. A tool that runs on the phone of the person at risk, rather than inside a carrier's network, could reach users that enterprise-only products miss.

What to watch is whether DetectifAI can show detection rates that hold up outside a demo. Real calls include background noise, multiple speakers and codec distortion, and a false positive that labels a real relative as fake could erode trust quickly. The company has not disclosed its accuracy figures or funding, so the Startup Battlefield appearance will be the first chance for outsiders to judge whether the on-device approach performs as promised.