Matt Welsh

Matt Welsh is Head of AI Systems at Palantir, co-founder of Fixie.ai, and a computer scientist whose career spans Harvard, Google, embedded machine learning, and AI application infrastructure. He argues that large language models are a computing platform: systems that execute human intent directly and change what programming itself requires.

Welsh studied computer science at Cornell and earned his doctorate at the University of California, Berkeley. As a Harvard professor, he researched operating systems, distributed systems, and wireless sensor networks, including volcano-monitoring deployments in Ecuador that collected seismic and acoustic measurements under severe connectivity and power constraints. Harvard granted him tenure and promoted him to full professor in 2010.

At Google, Welsh eventually directed Chrome Mobile engineering in Seattle, working on Chrome Data Saver, offline reading, and cloud-assisted optimization for people using limited devices or expensive, unreliable connections. In 2019, he joined Xnor.ai to develop machine learning for embedded hardware; after Apple acquired Xnor, he continued working on on-device intelligence. He subsequently led engineering at OctoML, building infrastructure for optimizing and deploying models across hardware platforms around the Apache TVM compiler ecosystem.

Welsh co-founded Fixie.ai in 2022, initially serving as chief executive and later as chief architect. Fixie combined a managed application platform with AI.JSX, an open-source TypeScript framework that gave frontend and full-stack developers composable building blocks for retrieval, model calls, external tools, conversational state, and deployment.

His AI Engineer Summit demonstration showed why the design exceeded a change in syntax: component trees could run several model calls concurrently, stream their outputs into downstream components, query indexed documents, and invoke GitHub’s GraphQL API. A voice-based doughnut-ordering assistant illustrated why low-latency conversational voice agents depend on responsive turn-taking as much as model quality.

• Models as computational infrastructure. In his essay on the future of programming, Welsh argues that models capable of translating goals into results could reduce the centrality of handwritten source code and shift engineering toward system design, evaluation, and specifying intent.
• Composable AI applications for web developers. AI.JSX applies familiar TypeScript and JSX patterns to parallel generation, streaming, retrieval, safety transformations, and API integration, making sophisticated AI behavior accessible within existing web applications.
• Engineering under real-world constraints. From volcano sensors and mobile data budgets to embedded inference and conversational latency, Welsh has repeatedly designed systems around bandwidth, hardware, responsiveness, and operational complexity.

After leaving Fixie in early 2024, Welsh became chief architect at Aryn, working on AI systems for complex, unstructured information. He subsequently joined Palantir to lead AI systems work focused on government, defense, and national security.