Malte Ubl is CTO of Vercel. He works on web infrastructure and AI-assisted software development, drawing on earlier responsibility for Google Search rendering and frontend systems. His recent writing focuses on feedback during development and the security boundaries around increasingly capable agents.[1][2]
Start with The Great Evals Debate — Ankur Goyal & Malte Ubl →
X Career
Ubl previously served as a principal engineer responsible for Google Search rendering and as an engineering director for Search. His retrospective describes creating the Wiz frontend framework and working with colleagues on shared serving infrastructure, AMP and page-experience improvements. He left Google on March 4, 2022, after more than eleven years. He also founded and curated JSConf EU, whose last edition in its original form took place in 2019.[1][3][10]
Contributions
In Everything hackable will get hacked, Ubl describes creating deepsec after asking whether AI code review could examine a whole codebase rather than just a proposed change. The open-source harness distributes review across isolated environments and produces vulnerability hypotheses for a person to investigate. His article argues for repeated review as models improve, while describing Vercel’s own use of the system.[4][5]
Ideas
Use several forms of feedback while building agents
In The Great Evals Debate, Ubl joins Ankur Goyal and swyx to discuss offline tests, production feedback and human judgment. The conversation examines how real user failures can become evaluation cases and how feedback supports faster iteration. Its focus is the relationship among those practices, rather than a single test score that settles product quality.[6]
Extract infrastructure from work the company already runs
His Ship AI conversation covers Vercel’s agent tooling, workflow primitives and Python support. The discussion connects those abstractions to the company’s own applications and agents, including v0. Ubl explains the infrastructure strategy as a way to support application developers; the products are the work of Vercel’s teams.[7]
Tools, writing & talks
Writing, research & talks
Start here
A discussion of offline evaluation, production failures and human feedback in coding-agent development.
Begin with the contrasting examples of offline tests and production feedback. The discussion is useful if you have tried a coding agent; use it to decide what feedback your own task needs before choosing an evaluation framework.
Participant with Ankur Goyal and swyx[6]Original source- Everything hackable will get hackedvercel.com
Ubl’s argument for ongoing defensive AI review and an account of creating deepsec.
Author[4]Original source A conversation about Vercel’s agent, workflow and deployment infrastructure.
Interview participant[7]Original source