Self-taught programmer Siqi Wen moved to Silicon Valley at age 17, living in a famous hacker house alongside notable artificial intelligence builders like Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown. During that period, Wen experimented with early software concepts that later shaped industry staples like Claude, Midjourney, and ChatGPT 3.5. Prominent investor Naval Ravikant even hired Wen to build Airchat during his early coding career. After learning how to run early open language models locally on consumer Apple hardware, Wen set out to build a completely private personal assistant.

Now supported by Peter Thiel’s Thiel Fellowship, Wen launched an invite-only beta for Underdog, positioning the platform as a privacy-focused competitor to mainstream assistants like Instinct and Muse. Underdog processes data entirely on local user devices, running on native macOS and Windows PCs with mobile versions for iOS and Android in active development. Keeping data on local hardware ensures private documents, chat histories, and personal files never leave the user’s laptop or smartphone.

Underdog relies on Husky, a custom inference engine built specifically to run language models efficiently on local consumer hardware. Husky speeds up execution times by minimizing data transfers between a computer’s central processor and its graphics card, enabling fast local processing on standard consumer laptops. To protect user privacy, Underdog encrypts access keys for connected email accounts and external services directly on the user device.

The software runs on a compact 2.7-billion parameter reasoning model fine-tuned from Qwen 2.5. Despite its smaller size compared to cloud-hosted models, the local software matches the performance of top frontier models from six months ago on standard benchmark evaluations. The local agent easily handles everyday personal tasks like drafting emails, managing schedule options, researching products, and helping students review homework problems. Small on-device models continue gaining reasoning speed, proving that consumers do not need to surrender personal data to access capable software assistants.

Underdog also introduces a unique business model that avoids traditional subscription fees and ad tracking. Because the software runs locally on consumer hardware, the startup avoids paying massive cloud server hosting bills. Instead of charging monthly fees, Underdog plans to process commercial transactions like ticket bookings through Stripe, taking a small payment fee similar to traditional credit card interchange rates. Aligning revenue with actual purchase transactions removes the incentive to monetize user data.

Traditional software platforms frequently collect detailed personal profiles, web browsing logs, and medical information to train models or sell targeted advertisements. In contrast, Underdog guarantees complete data privacy by keeping every piece of user information strictly on local drives. backed by major venture firms like Andreessen Horowitz, Founder Collective, and SV Angel, Underdog offers a compelling blueprint for privacy-focused consumer software. Running intelligent software locally on personal devices gives users full control over their private data without sacrificing daily productivity.