After a Deepfake Call Targeted Her Grandfather, This Founder Built On-Device Defenses

Preface

Context: A personal scam using a convincingly fabricated voice pushed Tarini Padmanabhuni to act. This article explains why on-device deepfake voice detection matters and how a startup translated a family crisis into a technical solution. The purpose is to outline the problem, the current market landscape, and the distinctive approach taken by the company she founded, DetectifAI. By emphasizing the need for real-time, private protection, the piece clarifies why device-level defenses differ from cloud-based solutions and how they can help people who lack tools to verify suspicious calls.

Lazy bag

Key takeaway: Deepfake voice scams are rising, hitting older adults hard. DetectifAI focuses on building compact AI models that run on phones to give instant detection without sending audio to the cloud. The startup targets phone makers first, offering an SDK that can be integrated as a system feature—and also licenses to businesses handling sensitive calls.

Main Body

The story begins with a distressing phone call. Tarini Padmanabhuni’s grandfather received a voice message he believed came from his brother, saying he had been kidnapped and demanding ransom. The imitation was convincing enough that he paid. Only later did the family discover the brother was safe and unaware; the voice had been created by AI. For Padmanabhuni, the incident was not primarily about the money lost, but about the victim’s absence of any reliable way to distinguish a real voice from a deepfake.

That experience became the catalyst for DetectifAI, a San Francisco startup built to detect AI-generated voices. The company’s founding premise is straightforward: as synthetic audio becomes more realistic and more accessible, people—especially older adults—are increasingly vulnerable to scams that exploit voice familiarity and trust. Law enforcement and industry data reinforce this concern. For example, recent figures show a significant uptick in AI-driven fraud losses, with millions of dollars lost and older demographics disproportionately affected.

Many companies are already working on deepfake detection, and the field includes established players and new entrants. Solutions exist from firms like Reality Defender, Pindrop, Resemble AI, Microsoft Azure AI Content Safety, and Nuance. These offerings, however, typically analyze audio in the cloud. While cloud-based detection can leverage large models and extensive compute resources, it also has limitations: latency in providing results, dependence on network connectivity, and privacy concerns because the audio leaves the user’s device. DetectifAI argues that these limitations matter especially in phone call scenarios, where immediacy and user privacy are paramount.

DetectifAI’s approach diverges by focusing on building compact models from the outset—models small enough to run within a smartphone’s operating system. Rather than taking large cloud models and attempting to compress them to fit on-device, the company designs lightweight architectures meant for mobile deployment. The advantage is twofold: detection becomes instant and private, and phone manufacturers can ship the feature as a built-in capability rather than relying on a third-party cloud service.

The company’s go-to-market strategy prioritizes phone makers. DetectifAI licenses a software development kit (SDK) that OEMs can integrate into their operating system, making deepfake detection a visible, differentiating feature of the handset—akin to how camera specs became a device selling point. Padmanabhuni suggests the first manufacturers to adopt on-device deepfake detection could gain competitive advantage, and she envisions deepfake detection becoming a standard specification in future smartphones.

Beyond device OEMs, DetectifAI plans secondary revenue through licensing to businesses that need to vet voice interactions—banks, collections agencies, and fraud-prevention firms among them. The startup already reports early traction: it processes more than 100,000 calls per month for financial institutions in India. In that setting, AI voice agents place calls for tasks like debt collection and loan follow-ups, and DetectifAI provides detection and speaker verification on every interaction to reduce the risk of misuse and impersonation. While Padmanabhuni declined to name clients due to confidentiality, the volume indicates both demand and operational capacity.

Padmanabhuni’s background helps explain her confidence. She started working with machine learning at a young age and studied cyber-physical systems at Manipal Institute of Technology. She also led a student engineering team building a driverless racecar—an early sign of combining software, hardware, and rigorous systems thinking. That mix of experience is relevant: building compact, reliable AI for phones requires an understanding of model design, systems constraints, and real-world deployment challenges.

Practical user testing has also yielded encouraging signs. A WhatsApp beta allowed users to forward suspicious voice notes and receive an assessment. One tester, whose relatives had been scammed, said they would readily pay for such a service—underscoring unmet demand. Padmanabhuni frames these early user reactions against her grandfather’s experience: the key difference is the presence of an accessible, reliable tool to judge whether a voice is authentic.

From a business perspective, DetectifAI has raised a small seed round from angel investors and early backers, giving it runway to refine technology and expand integration discussions with handset makers. The startup is also participating in high-visibility industry programs, which can accelerate partnerships and market recognition. As deepfake audio becomes a mainstream security and privacy concern, companies that can combine efficacy, speed, and privacy—especially on-device—may find receptive partners across hardware and enterprise markets.

However, challenges remain. On-device models are constrained by compute, memory, and battery limitations, which can restrict the complexity of detection architectures. Adversaries continuously refine synthesis techniques, requiring models to adapt and stay current. Integration with diverse handset platforms and regional regulatory environments will also influence adoption timelines. Finally, user education and clear UX are necessary so people understand what a detection result means and how to act on it.

Ultimately, DetectifAI’s proposition rests on a clear trade-off: accept smaller, specialized models that protect privacy and provide instant feedback, or rely on large cloud models that may offer greater raw detection capability but at the cost of latency and privacy. For use cases like phone calls—where immediacy and trust are critical—on-device detection presents a compelling solution. For Padmanabhuni, the mission remains personal: to give others the ability her grandfather lacked, a straightforward way to know whether the voice on the line is real or fabricated.

Key Insights Table