Underdog Launches as a Privacy-Focused On-Device AI Assistant from Silicon Valley Prodigy

Image: TechCrunch · Source

Sigil Wen, backed by major investors, has launched Underdog, an on-device AI assistant designed to maximize user privacy while delivering everyday AI tasks without data leaving users' devices.

Self-taught programmer Sigil Wen, known in Silicon Valley for his early work on prominent AI projects, has unveiled Underdog, an AI assistant promising unprecedented user privacy by running entirely on personal devices. Wen’s experience includes collaborations with renowned AI researchers and contributions to foundational tools that evolved into today's popular AI systems. Now, as a Thiel Fellow, he has released an invite-only beta of Underdog, initially available on Macs and Windows PCs, with plans for Linux, iPhone, and Android versions.

Underdog is powered by Husky, a custom inference engine developed by Wen to efficiently run AI models on local hardware. This design minimizes data transfer between a computer’s main processor and graphics chip, contrasting with other on-device engines. Enhanced security features include encryption of keys for email and other accounts authorized by users.

Unlike many AI assistants relying on large cloud-based models, Underdog currently operates a 27-billion parameter reasoning model fine-tuned from Qwen3.8-27B. Wen asserts this model delivers performance on par with benchmarks achieved by larger models such as Claude Opus 4.6 from half a year ago, suitable for tasks like shopping research or math help. He emphasizes that users do not have to compromise privacy to gain AI capability.

The project’s business approach is notable: Underdog is free and will remain ad-free, enabled by low operational costs since the AI runs on users' devices without extensive cloud inference expenses. Supported by investors including Stripe co-founder Patrick Collison and Andreessen Horowitz, Underdog's revenue model draws from taking a small percentage of payments processed through Stripe when transactions occur via the assistant. This approach aligns the AI’s incentives with users rather than relying on data monetization strategies common in other services.

Wen highlights the risks of data collection by AI assistants, which often access sensitive personal information, and advocates for privacy as a fundamental right when using AI. He describes Underdog as a product he would trust for his own family, aiming to build an assistant that respects privacy without sacrificing functionality.

Conway Research, the startup behind Underdog, counts high-profile venture and angel investors, underscoring strong confidence in this privacy-centered AI assistant concept. The launch represents a significant step toward AI tools that empower users while safeguarding their personal data.