Privacy in AI has long felt like a broken promise. You hand over your most sensitive data, medical history, financial details, even information about your kids, and in return you get a smarter chatbot that quietly feeds that data back into the machine. A new entrant called Underdog is betting that model is both ethically wrong and commercially unnecessary.
A Founder Shaped by AI’s Earliest Days
Sigil Wen is not your typical startup founder. A self-taught coder who moved to Silicon Valley at 17, he lived and coded alongside people who would go on to shape the AI industry as we know it. He tested early prototypes of what became Claude, Midjourney, GPT-3, and Stable Diffusion before most of the world had heard of any of them. That firsthand exposure to the raw potential of AI, combined with growing concern about where user data actually goes, became the foundation for Underdog.
Now a Thiel Fellow, Wen has launched an invite-only beta of Underdog through his startup Conway Research. The app runs entirely on-device, currently supporting Macs and Windows PCs, with Linux, iPhone, and Android support planned. At its core is Husky, a custom inference engine Wen built to run AI models faster by reducing data movement between a machine’s main processor and its graphics chip. That technical edge matters because on-device AI has historically suffered from sluggish performance.
Smaller Models, Smarter Trade-offs
Underdog currently runs a 27-billion parameter reasoning model fine-tuned from Qwen3.8-27B. That is considerably smaller than the cloud-hosted giants powering competing assistants, but Wen argues the gap is narrowing fast. He claims performance is comparable to Claude Opus 4.6 on certain benchmarks, roughly what was considered state-of-the-art just six months ago. For everyday tasks like shopping research, scheduling, or answering questions, that is more than enough.
The privacy architecture goes beyond simply keeping data local. Underdog encrypts the keys to any email or external accounts users authorize it to access. No data leaves the device, no ads are served, and no user profiles are sold. That is a meaningful structural departure from virtually every mainstream AI assistant available today, whose terms of service typically allow broad data collection and use in model training.
A Business Model Built Around Alignment, Not Extraction
Perhaps the most disruptive element of Underdog is how it plans to make money. Because the AI runs on users’ own hardware, Wen avoids the massive inference costs that force competitors to charge monthly subscriptions or monetize data. Instead, Conway Research will take a small percentage of payment transactions the assistant facilitates, using Stripe’s infrastructure. Stripe co-founder Patrick Collison is among the angel investors backing the company, alongside Andreessen Horowitz, Khosla Ventures, and others.
For consumers who are actively evaluating AI assistants and wondering which one to actually trust with sensitive tasks, Underdog represents a genuinely different category. If on-device AI continues improving at its current pace, the buying decision for privacy-conscious users could shift dramatically within the next 12 to 18 months. Underdog is not just a product launch. It is an early signal of where the market is heading.
