A self-taught coder who moved to Silicon Valley at 17 has introduced an invite-only AI assistant designed to operate without transmitting user data to external servers. Sigil Wen, a Thiel Fellow and former collaborator with early AI researchers including Andrej Karpathy, launched Underdog on Monday. The assistant runs entirely on users' personal devices — currently Macs and Windows PCs — with versions for Linux, iPhone, and Android planned.

Wen developed Husky, an inference engine that minimizes data movement between a computer's processor and graphics chip, to enable on-device operation. Underdog uses a 27-billion parameter reasoning model fine-tuned from Qwen3.8-27B, which Wen states performs comparably to Claude Opus 4.6 in certain benchmarks for everyday tasks like answering questions or researching purchases. The model's capabilities, he argues, match those of larger cloud-based systems for common user needs.

The system incorporates privacy safeguards, including encryption of authentication keys for linked accounts. Unlike many AI assistants that collect user data for advertising or model training, Underdog does not monetize personal information. Instead, it plans to generate revenue through a transaction fee structure using Stripe's payment infrastructure, taking a small percentage of payments processed via the assistant. This approach eliminates the need for subscription fees or data harvesting.

Wen described the project as one he would personally use and trust for his family, emphasizing that privacy should not require sacrificing functionality. The startup behind Underdog, Conway Research, has secured funding from investors including Andreessen Horowitz, Khosla Ventures, and Stripe co-founder Patrick Collison. The project remains in beta testing, with no public release date announced.