DetectifAI Wants Your Phone to Spot a Cloned Voice Live
DetectifAI, a San Francisco startup founded by Tarini Padmanabhuni, is building compact AI models that live inside a phone operating system and give an instant verdict on AI-generated voice. TechCrunch
The company was founded in 2025. Its core product is a software development kit, a toolkit other companies can build into their products, first licensed to phone makers as a built-in operating system feature. YourStory
A personal starting point
The origin was personal. About two years before September 28, 2026, Padmanabhuni's grandfather paid a ransom after a call in which a cloned voice imitating his brother said the brother had been kidnapped. The voice was synthetic.
That type of fraud has scaled quickly. Americans lost close to $900 million to AI-driven scams last year, up 24% from 2024, according to FBI data cited in the article. People age 60 and older lost twice as much as those aged 50 to 59.
How the technology works
The technical bet is on-device inference, meaning the check runs on the phone itself. Its models are meant to work during calls, voice messages and other audio without audio leaving the device. The case is latency, or speed, and privacy. A verdict has to arrive while the call is still live, and keeping raw audio on the handset avoids the round trip to a cloud classifier, a large checker running on distant servers.
The business plan so far
The go-to-market starts with the OS layer. DetectifAI is first selling to phone manufacturers by licensing software tools for detection as a built-in phone operating system feature. Secondary revenue is planned from licensing the technology to businesses and fraud-prevention firms.
It already has early revenue outside handsets. DetectifAI handles more than 100,000 calls per month for financial institutions in India. Those calls are placed by AI voice agents handling debt collections and loan-document follow-ups, with deepfake detection and speaker verification on every call.
The company also ran a small WhatsApp beta test where users forward suspicious voice notes and receive an assessment of whether they are real. It raised a small seed amount including investment from Josh Constine, formerly an editor at TechCrunch.
Padmanabhuni says she began working in machine learning at age 12. She studied cyber-physical systems at Manipal Institute of Technology in India, and says she became the youngest team lead of India's first driverless racecar division in Formula Student.
Why this is hard to build
The broader context here is familiar to mobile audio work. Call-audio access is constrained, DSP and NPU budgets, the processing set aside for sound and AI tasks, are tight, and telephony codecs discard much of the spectral detail that detectors use. A compact model that survives AMR, Opus and speakerphone distortion is a different engineering problem from a large classifier scoring clean files in the cloud.
For buyers, the two tracks complement each other. A kit for phone-maker integration covers the consumer scam case, where the victim hears the cloned voice directly. The India deployment covers the inverse case, where the enterprise places automated calls and must confirm the human on the other end is live and legitimate. Both require low false-positive rates. A detector that cries wolf will be disabled.
In my view, the upkeep load is worth flagging. Voice synthesis quality moves fast, and detection features decay. On-device distribution through phone-maker updates can help, but it also ties model refresh to someone else's release train. The WhatsApp forwarding test gathers wild-type samples and user tolerance data without waiting for OS deals to close.
If this approach holds, the payoff is straightforward. A phone that can flag a synthetic voice in the audio path, privately and in real time, gives older users and high-volume lenders the same basic protection that spam filtering once gave email. The profile was published September 28, 2026, ahead of TechCrunch Disrupt 2026, scheduled for October 13-15 at Moscone West in San Francisco. TechCrunch
