Underdog AI Assistant Launches With a Privacy-First Approach and On-Device Intelligence
Sigil Wen’s journey into artificial intelligence started long before most people were talking about AI assistants, local language models, or privacy-focused chatbots. As a self-taught coder, Wen moved to Silicon Valley at just 17 and found himself living in an AI hacker house alongside influential researcher Andrej Karpathy.
That environment placed him close to some of the people and ideas that would later help shape the modern AI boom. He worked and experimented around future leaders in the field, including Perplexity founder Aravind Srinivas and OpenAI researcher Noam Brown. He also got early exposure to AI systems that would eventually evolve into major products, including early chatbot, image generation, and large language model projects.
At one point, Wen even managed to get GPT-2 running on an Apple Watch, simply for fun. For him, that period represented a rare moment of creative energy, when many of today’s biggest AI ideas were still experimental projects being tested by a small group of deeply curious builders.
Now, Wen is turning that experience into a new product called Underdog, an invite-only beta AI assistant designed around one central promise: keeping user data private.
Unlike many popular AI assistants that rely on cloud servers to process requests, Underdog runs entirely on the user’s own device. At launch, it supports Mac and Windows PCs, with Linux, iPhone, and Android versions planned for the future. Because the AI model runs locally, personal information stays on the hardware users already own rather than being sent to remote data centers.
The technology behind Underdog is powered by Husky, an inference engine built by Wen to run AI models quickly and efficiently. Inference engines are the software systems that allow AI models to process prompts and generate responses. Husky is designed to reduce the amount of data moving between a computer’s main processor and graphics chip, helping the assistant perform faster on consumer devices.
Privacy is not limited to local processing. Underdog also includes additional security measures, such as encrypting account keys for services like email and other connected tools that users choose to authorize. This approach is meant to give the assistant useful access without exposing sensitive credentials unnecessarily.
Underdog currently uses a 27-billion-parameter reasoning model fine-tuned from Qwen 3.8 27B. While this is much smaller than the largest frontier AI models running in massive data centers, Wen believes it is powerful enough for many everyday tasks. He says the model can handle activities such as product research, answering homework-style math questions, organizing information, and helping users complete common digital tasks.
His argument is simple: most people do not need to give up their privacy to get a capable AI assistant.
“You don’t need to sacrifice your privacy for the capability because they’re just as capable,” Wen said, adding that smaller on-device AI models are likely to become even more useful over time.
One of the most unusual parts of Underdog is its business model. The app will be free at first and will not rely on ads. Since the AI runs on users’ own devices, the company does not face the same high cloud-computing costs that many AI startups must cover through subscriptions.
Instead of charging users to run the assistant, Underdog plans to earn a small percentage from payment transactions the AI completes through secure payment infrastructure. The idea resembles a small transaction fee, similar to how parts of the financial technology industry operate.
This model is meant to avoid one of the biggest concerns surrounding AI assistants: data monetization. Many AI products depend on collecting user data, using it to improve models, support advertising, or share insights with third parties. That trade-off becomes especially sensitive when an AI assistant is expected to help with deeply personal areas of life, including finances, health, family information, schedules, shopping, and private communication.
Wen has made privacy a core part of Underdog’s identity. In his AI manifesto, he raises a direct question: why should using AI require surrendering private information?
He has also said that he is building Underdog as a product he would personally trust and feel comfortable letting his future children use.
The startup behind Underdog is called Conway Research, and it has attracted backing from major investors. Supporters include Andreessen Horowitz through partner Chris Dixon, Khosla Ventures, Hummingbird, SV Angel, the Anthology Fund, and several notable angel investors. Stripe co-founder Patrick Collison is also an investor, along with Vercel founder Guillermo Rauch, Noam Brown, Deedy Das, and others.
Underdog enters the market at a time when AI assistants are becoming more powerful, more personal, and more integrated into everyday life. That makes the question of privacy increasingly important. If an assistant is expected to manage tasks, understand preferences, search through personal accounts, and potentially make purchases, users may want stronger guarantees that their information is not being stored, analyzed, or sold elsewhere.
By focusing on local AI processing, encrypted access, and a non-advertising business model, Underdog is positioning itself as a different kind of AI assistant. Rather than asking users to trust remote servers with their most sensitive data, it aims to bring AI capability directly onto personal devices.
The challenge will be whether Underdog can deliver the speed, intelligence, and convenience people expect from leading AI tools while keeping everything private and on-device. If it succeeds, it could become an important example of where consumer AI may be heading: powerful, personal, and far less dependent on the cloud.
