After a deepfake voice fooled her grandfather, this founder sprang into action

We have reached the point where the cost of entry for a sophisticated heist is essentially zero. A few years ago, if you wanted to impersonate a family member to steal life savings, you needed a professional actor and a lot of luck. Today, you just need a ten-second clip from a social media video and a subscription to a cloning service. It is a massive problem for the average person, but for founders, it represents one of the most urgent frontiers in security.

The Personal Cost of Synthetic Media

Tarini Padmanabhuni experienced this shift firsthand, not through a research paper, but through her grandfather. A scammer used an AI-generated clone of his brother's voice to manipulate him. It worked because the emotional response to a familiar voice bypasses our logical defenses. When your brother calls you in a panic, you do not stop to analyze the metadata of the call.

This event led to the creation of DetectifAI, a San Francisco-based startup that just landed in the spotlight at TechCrunch Disrupt. While many companies are trying to solve deepfakes by building massive databases of known signatures, Padmanabhuni is taking a different route: local, real-time detection on the device itself. This is a builder-first approach that respects the reality of how we use technology.

The Technical Barrier: Latency vs. Accuracy

If you are a builder in the AI space, you know the trade-offs between model size and performance. To catch a deepfake during a live conversation, you cannot afford to send the audio to a remote server, process it, and send an alert back. By the time the cloud says the voice is fake, the victim has already shared their bank details. The latency kills the utility.

DetectifAI is focused on shrinking these models so they can run directly on a smartphone. This is the hardest way to build, but it is the only way that actually works for the end user. It requires optimizing the math so it does not drain the battery or lag the call, all while maintaining a high enough confidence interval to avoid false positives. If the app flags your actual grandmother as a robot, the user will delete the app in a heartbeat.

Why Local Processing Matters

There is also a massive privacy angle here that founders often overlook in their rush to move fast. If you build a tool that monitors phone calls for fraud, but that tool requires uploading private conversations to a company server, you have created a new security nightmare. You are essentially asking users to install wiretapping software in exchange for safety.

By keeping the processing on the device, DetectifAI avoids this trap. The data never leaves the phone. The model listens for the artifacts and synthetic patterns that characterize AI speech—things like unnatural breathing patterns or the specific robotic cadence that human ears miss but algorithms can spot—and flags them locally. This is the standard we should be holding AI safety tools to.

The Reality of the Arms Race

I am usually skeptical of "AI for good" pitches because they often lack a sustainable business model or a technical moat. However, the market for voice authentication and verification is exploding. We are moving toward a world where voice is the primary interface for everything from smart homes to banking. If the voice cannot be trusted, the entire ecosystem collapses.

The challenge for Padmanabhuni and her team is that the scammers are also using AI to improve. As generative models get better at mimicking human imperfections, the detection models have to get more sensitive. It is a permanent arms race. For builders, the lesson here is that you cannot just build a product and walk away. A security startup in the AI age is a service that requires constant iteration.

What Builders Should Take Away

• Focus on the Edge: The future of AI utility is local. If your tool can run on a device without a massive GPU cloud, you have a competitive advantage in both speed and privacy.
• Emotional Solving: The best products often come from solving a personal pain point. Padmanabhuni didn't just see a market gap; she saw a family member get hurt. That clarity of purpose usually leads to better product-market fit.
• Avoid Data Bloat: You do not always need more data; you need better optimization. Shrinking a model to fit a mobile chip is a massive technical hurdle, but it creates a much higher barrier to entry for competitors.

The Skeptics Corner

Is a mobile app enough? Probably not. A determined attacker will always find a workaround, and as long as there is a human on the other end of the line, there is a vulnerability. Detection software is a layer of defense, not a silver bullet. We also have to consider the "crying wolf" effect. If every other call triggers a warning because of poor signal quality or a heavy accent, people will stop paying attention.

However, the work being done at DetectifAI is a necessary step. We cannot rely on the platforms like Apple or Google to solve this for us quickly enough. They move slow; startups move fast. The transition from "I think this sounds like my brother" to "My phone tells me this is 99% likely to be a computer" is the shift we need to protect vulnerable populations.

The most dangerous thing about AI isn't that it will become sentient; it's that it makes it incredibly cheap to be dishonest at scale.

We are watching the birth of a new sector in cybersecurity. It is no longer just about firewalls and passwords; it is about verifying the biological reality of the person on the other side of the screen. For builders, this is one of the few areas where you can build something that actually matters while the rest of the world is busy making cat memes and wrappers for ChatGPT.

Read the original at TechCrunch AI →