The interview began with a broken cluster. Greg Brockman had tried to set up Kubernetes, which in late 2015 was young enough to make every decision feel faintly experimental. Could Vicki Cheung fix it? She had never used Kubernetes. She had not used Terraform either. There was no formal interview process, because the organization that would become OpenAI was not yet properly an organization. Cheung and Brockman were working from his apartment. There was no office waiting, and no company laptop. There was merely a problem, two people, and the useful possibility that the problem might be fun.
Cheung began pulling at the network. She used packet captures to find out why the machines were not speaking. By the end, something worked, and she was hired. The episode has the tidy dimensions of a founder fable, but its appeal lies in the untidiness. Nobody knew exactly what the job was. The tool was unfamiliar. The infrastructure had to exist before researchers arrived to use it. Cheung's career has repeatedly started in that hour before the furniture turns up.
“I actually really see myself as a generalist because I find all things engineering very fascinating.”Vicki Cheung
Her own summary is terser: “I build things.” It appears on her personal site above a record that runs from Duolingo and TrueVault through OpenAI, Lyft, Gantry, and now Harmonic, where current professional profiles place her as a vice president. These names tempt a history of prestige. The more revealing history is one of layers. Cheung began near the screen, moved through mobile and APIs, descended into the machinery underneath, then returned to the human beings whose work that machinery was meant to improve.
The internet did not ask her age
Cheung grew up in Hong Kong and met computing at about ten, when public schools were beginning to offer computer classes. The equipment was plain. The lesson included Logo and HTML, written by hand because there was no luxurious software to hide the tags. What caught her was not merely the puzzle. A ten-year-old could publish a site on GeoCities and the visitor would not know a child had made it. Code offered the small, thrilling deceit of competence: the work could arrive before the biography.
She followed that permission into PHP sites and, in high school, college computer-science textbooks that she read and worked through for pleasure. At a Carnegie Mellon summer program, somebody suggested that she stay. So she left high school and began college. It sounds reckless only if one ignores the preparation already under way in her bedroom and browser.
At Carnegie Mellon she approached professors whose work interested her and asked to help. Luis von Ahn had a project with his doctoral student Severin Hacker. It was called Duolingo. On Cheung's first day, a large natural-language-processing book sat on her desk. The assignment was to learn about speech recognition and work on the engine. Her response, in later recollection, was essentially: sure, that sounds fun.
Doing whatever the product required
Duolingo grew from four people to forty during Cheung's time there. The work refused to stay in one lane. After speech recognition, she built much of the web front end, learned Android on the job and produced early versions of the app, worked on APIs, localization and experimentation, and led the Incubator that let contributors create language courses. She tested the Spanish course so repeatedly that, years later, a podcast host asked for a shortcut after his progress had decayed. “There's no cheating out of learning,” she told him. “The algorithm thinks you've forgotten.”
The joke contains a philosophy. Systems are allowed to be friendly, but they cannot negotiate away reality. Learning takes repetition. Reliable software takes an acquaintance with failure. Early-company engineering takes a willingness to handle the unglamorous task that appeared this morning. Cheung drifted lower through the stack not because the user interface had become unworthy, but because each layer exposed the constraint beneath it.
The build path
TrueVault followed, with core API and infrastructure work. Then came the apartment, the cluster and OpenAI. Deep-learning research did not behave like the stateless web services Kubernetes had been designed to host. Researchers launched irregular batches of expensive experiments. Jobs could occupy unusual combinations of processors and graphics hardware. Cheung and her colleagues built around those mismatches, including a batch-oriented autoscaler, while the broader ecosystem was still deciding what Kubernetes ought to become.
The expensive mistake of knowing the user
The most important correction arrived without a new piece of software. Cheung's infrastructure group built tools it assumed researchers wanted. The assumptions were reasonable. Researchers were technical; infrastructure engineers were technical. Surely one technical person could anticipate another. They could not.
So Cheung sat beside the researchers and watched them work. She tightened the feedback loop. Instead of replacing an entire workflow with a magnificent new regime, the team made smaller changes that accommodated habits already carrying useful knowledge. Her later advice was compact: “Definitely assume less and listen more.” It is unusually humane counsel from the engine room.
“If you don't know what the person you're working with is optimizing for, it's very hard to work effectively as a team.”Vicki Cheung
At Lyft, where she joined in 2018, the scale changed again. The company already ran hundreds of services on AWS, none of its production workload yet containerized. Leading the compute team, Cheung helped begin with batch and machine-learning jobs, whose episodic nature made them easier to move than services continuously handling traffic. Her infrastructure experience traveled well precisely because the companies did not resemble one another. Everyone, she observed, needs infrastructure simply to get started.
Management expanded the definition of infrastructure. Cheung held regular conversations about what team members wanted from their careers and looked for ways the company could supply it. She also accepted that sometimes the right answer was to leave. A manager's system includes the working environment, the possibility of growth, and who feels able to remain in the industry. Architecture diagrams rarely draw those boxes, which does not make them optional.
A feedback loop for machines in the wild
By 2020, Cheung had seen the same gap from several directions. Conventional software engineering had developed mature ways to observe running systems, deploy changes and respond when reality disagreed with a plan. Machine-learning teams faced a stranger bargain. A model could perform handsomely in training and then meet users, changing behavior, new language, or an unforeseen edge case. The world would move while the model remained still.
She co-founded Gantry with former OpenAI researcher Josh Tobin to help teams understand how machine-learning applications behaved after deployment and feed what they learned into the next version. Gantry emerged from stealth in 2022 with $28.3 million in announced seed and Series A financing. Its proposition was less theatrical than artificial intelligence often prefers: measure the live product, find the weak slices, collect human feedback, improve it, repeat.
Build the first useful version. Watch the real workflow. Find where the system resists its users. Shorten the distance between failure and learning.
The thread from GeoCities to Gantry is not a specific language, framework or title. It is agency made accountable by observation. Cheung likes the moment when a person realizes they can make a thing. She also knows that making it once is insufficient. Products acquire users. Research acquires scale. Models acquire a life outside the dataset. Each success creates a new underside that somebody must understand.
Her public profile is quieter after Gantry, whose tenure records end in 2024. In 2026, LinkedIn identifies Harmonic as her current organization, and a professional data profile lists her there as a vice president. The available record does not yet offer the richly narrated chapter that her earlier interviews do. The restraint is fitting. Cheung has seldom sounded interested in a career that reads neatly while it is happening.
No twenty-year architecture
Asked where she hoped to be in twenty years, Cheung declined the false precision of a roadmap. Twenty years earlier, she could not have predicted the path from handwritten HTML in Hong Kong to an AI laboratory in San Francisco. She offered wishes instead. She wanted to keep applying engineering to problems she cared deeply about. She wanted still to be in technology, and to help make it a better place for women who enter and too often find reasons to leave.
There is no grand claim in that answer, and no fondness for inevitability. Cheung credits luck, timing and the active pursuit of work she enjoyed. Yet luck has met a reliable accomplice: when confronted with an unfamiliar system, she tends to look closer. When the system fails, she traces the conversation. When the elegant tool annoys its user, she pulls up a chair.
The infrastructure is therefore more than clusters. It is the arrangement that lets another person proceed. A child publishes a page. A learner hears a phrase. A researcher launches an experiment. An engineer understands what a live model is doing. The best foundations disappear beneath the thing they enable. Cheung's career is visible because so much of her work was designed, quite deliberately, not to be.