Sigil Wen, a Thiel Fellow with deep roots in Silicon Valley’s elite AI developer circles, has officially launched the invite-only beta for Underdog. Developed by Wen’s startup, Conway Research, Underdog is a personal AI assistant designed to run entirely on a user's own hardware. The launch represents a direct challenge to the data-collection models of current industry leaders, positioning privacy not as an add-on, but as the core architecture of the product.
Wen is a well-known figure in the San Francisco AI scene despite his young age. At 17, he moved to the valley and lived in a hacker house alongside prominent figures like Andrej Karpathy and Perplexity founder Aravind Srinivas. During this period, he was an early tester for technologies that would eventually become Midjourney, Claude, and GPT-3. After a stint at Naval Ravikant’s Airchat, Wen is now focusing on what he calls his "AI manifesto," which questions why users must surrender their private information to access high-level intelligence.
Localized intelligence via the Husky engine
The primary differentiator for Underdog is its local execution. Unlike popular assistants that send queries to massive data centers, Underdog runs its models locally on Macs and Windows PCs. Linux, iPhone, and Android versions are currently in development.
To make this possible without draining system resources or suffering from significant lag, Wen developed a proprietary inference engine called Husky. According to Wen, Husky is designed to run AI models more efficiently by reducing the amount of data moved between a computer’s central processor and its graphics chip. This optimization is intended to provide a snappy, responsive experience even when the assistant is handling complex tasks on consumer-grade hardware.
By keeping the computation on-device, Underdog ensures that sensitive user data, ranging from personal emails to financial records, never leaves the owner's machine. The software also includes baked-in security features, such as the encryption of access keys for any third-party accounts a user authorizes the assistant to manage.
Performance and model architecture
While on-device AI is often associated with lower performance due to the lack of massive server farms, Underdog utilizes a 27-billion parameter reasoning model. This model is a fine-tuned version of Qwen3.8-27B.
Wen claims that this 27B model compares favorably against Claude Opus 4.6 in specific benchmarks, roughly matching the performance levels that were considered state-of-the-art only six months ago. He argues that for the vast majority of consumer needs, such as math assistance or shopping research, users no longer need to sacrifice privacy for capability.
"You don’t need to sacrifice your privacy for the capability because they’re just as capable," Wen said, noting that small, on-device models are rapidly narrowing the gap with their cloud-based counterparts.
A fintech-inspired business model
Perhaps the most unconventional aspect of Underdog is its monetization strategy. Most AI companies either charge a monthly subscription fee to cover high server costs or monetize user data through advertising. Because Underdog runs on the user's hardware, Conway Research does not face the massive monthly bills for cloud inference that plague other startups.
Supported by Stripe co-founder Patrick Collison, who is an angel investor in the company, Underdog is adopting a fintech-style revenue model. The app is free to use and will not feature advertisements. Instead, the company plans to earn revenue through a tiny percentage of payment transactions made through the assistant using Stripe’s secure infrastructure.
This "interchange fee" model aligns the company’s incentives with the user's financial utility rather than their data. By taking a small cut of commerce facilitated by the AI, the company can remain profitable without ever needing to sell user information to third parties or advertisers.
Solving the trust gap in AI
The launch of Underdog comes at a time when users are increasingly wary of how AI companies handle intimate details. Assistants are often most useful when they have access to a user's medical history, financial status, or family schedules. However, traditional cloud-based privacy policies frequently allow companies to use that data for model training or advertising.
Wen told TechCrunch that he wanted to build a product that he would be proud to have his own children use. His manifesto for the project centers on the idea that personal AI should be as private as a local file on a hard drive.
The startup has attracted a significant roster of high-profile investors. In addition to Patrick Collison, Conway Research is backed by Andreessen Horowitz via partner Chris Dixon, as well as Khosla Ventures, Hummingbird, and SV Angel. The company also received backing from the Anthology Fund, a partnership between Menlo Ventures and Anthropic. Other notable angel investors include Vercel founder Guillermo Rauch and OpenAI researcher Noam Brown.
What happens next?
Underdog is currently in an invite-only beta phase for desktop users. The company is working to expand its footprint to mobile devices, which will be a critical test for the Husky inference engine's efficiency on battery-powered hardware.
As the industry moves toward more agentic AI, software that can perform actions like booking flights or buying groceries, the security of those transactions becomes paramount. Underdog’s bet is that users will prefer an assistant that is "aligned with you as your bank," rather than one that treats personal data as a secondary product.
Filed under: AI, TechNews, Startups, Software, ProductLaunches, Privacy