Serena Ge keeps changing the assignment

The revelation arrived with an egg crepe and a grape-flavoured Hi-Chew. Serena Ge had been working on an app, thinking about the thousands of people software could reach. Downstairs, her mother was making breakfast. Her little sister handed her a sweet. The people closest to her had become background scenery while she pursued a grander idea of helping people.

She later described the episode in an essay about chasing impact. Users had become database entries; progress meant numbers moving on a dashboard. Breakfast offered a less convenient measure. How much attention was she giving the people who actually knew her? A confectionery company had, inadvertently, supplied a fairly effective management consultant.

Keep that scene in mind when considering her next business. Ge is the co-founder and CEO of Datacurve, which supplies coding data for training and evaluating AI models. Its work depends on persuading capable people to do difficult things well. The machines get the attention. Her business begins with the humans.

A training plan of her own

Before coding models, there was climbing. In high school, Ge wanted personalised training she could afford. She built an app and worked with athletes from Team Canada and the World Cup circuit. The project joined two activities that would keep reappearing in her life: making software and testing whether it helped someone accomplish something concrete.

She qualified twice for Canadian bouldering nationals. That achievement gives the app a pleasingly demanding customer reference: its maker had to use it herself. Her LinkedIn profile also records a first-place finish in the junior girls category at a York Region bouldering competition. These were actual competitions, with results beyond a screen.

Climbing makes a useful introduction to Ge without requiring every business decision to become a mountaineering metaphor. She had a problem, built a tool, and put it to work. The experience helped draw her towards computer science at the University of Waterloo. Software had already escaped the category of homework.

The sixth course was philosophy

At university, Ge was interested in more than getting technically employable. In January 2023, she wrote about adding a sixth course so she could attend philosophy classes at Wilfrid Laurier. Her friends wanted an explanation. The extra work was intentional: she wanted room to think about assumptions that a programming assignment might never ask her to examine.

She valued the structured discussions and the chance to develop ideas with other people. Her argument for humanities courses was personal, rather than a promise that philosophy would improve anyone’s salary. A university could cultivate the person doing the work as well as the skills that made the work possible.

“You are a person before you are any profession.”

Serena Ge, January 2023

That line complicates the familiar dropout story. Ge left a computer science programme, but she had also voluntarily made her university workload larger. Her complaint was about the boundaries of an education organised around employment. The attraction of a startup was partly the opportunity to choose the question, instead of simply completing the assigned answer.

UncleGPT did not survive the journey

Ge met Charley Lee through Waterloo’s AI reading groups and advanced computer science classes. They began building together. One project was UncleGPT, a planning tool that used an AI agent to break goals into tasks. Their early account recorded 930 users, roughly 1,300 projects and 5,500 messages. People tried it. The founders nevertheless moved on.

During a 2023 internship at Cohere, Ge worked on improving language-model reasoning. That experience brought her close to a less photogenic part of the AI industry: the difficulty of getting suitable training material. A model’s capabilities depended on the examples available to teach it. Finding good coding examples was a practical obstacle.

She and Lee entered Y Combinator’s Winter 2024 batch and changed direction three times before arriving at Datacurve. A tidy founder biography can make a company sound inevitable. Three pivots in a single accelerator batch suggest a more interesting process: build, discover the problem with what you built, and change your mind while there is still time.

Near the end of the batch, Ge wrote about learning to handle unfamiliar responsibilities. She could identify what she needed to learn, find it out, and act. Credentials were less useful than that sequence. It was an unusually practical answer to the question of whether someone was ready to run a company.

Give the engineers a reason to stay

Datacurve’s proposition starts with a recruitment difficulty. Skilled programmers already have things to do. Asking them to become data annotators is unlikely to sound like an irresistible evening. Yet an AI company that needs examples of difficult software work needs precisely the judgment those programmers possess.

Ge’s answer was to make participation appealing on its own terms. Datacurve built a gamified platform where engineers solve coding challenges for rewards. Competitive programmers and experienced developers could contribute work that became training data. Challenge, community and payment all had a place in the arrangement.

By October 2025, the company had distributed more than $1 million in bounties. That figure represents payments to contributors, rather than company revenue or funding. It is useful because it shows the collection system operating beyond an idea on a pitch slide. Engineers had done work, and the platform had paid them.

Onboarding and the working experience become more consequential as tasks get longer. Ge has described the need for guidance and validation throughout complex assignments. A willing contributor still needs tools that make it possible to finish. Paying someone to begin is only part of the job.

Datacurve’s current products extend that concern to expert demonstrations, execution traces and tasks that stretch over hours or days. A finished answer hides many of the decisions that produced it. Checks, failed approaches and recoveries can contain useful information too. The commercial task is to capture that work in a form a model developer can use.

The apartment came with futons

Ge’s early picture of San Francisco entrepreneurship owed something to HBO’s Silicon Valley. In a conversation on the Pre Seed Show, she described an apartment in Dogpatch shared by the founders and early teammates, with Japanese futons on the floor. Television had apparently left an impression on the accommodation policy.

The detail is funny; the explanation behind it is more revealing. Ge spoke about wanting a lasting bond among the people building the company. She saw ambition and something to prove as advantages the young team could bring. Living and working together was part of that early experiment in making a company feel like a shared project.

Her own reflections at the end of YC return to the pleasure of building with friends. She also described a cluster of setbacks shortly before demo day that forced her to consider whether she would prefer school or an engineering job. She still wanted the uncertainty of making a business. That preference is more specific than a general enthusiasm for entrepreneurship.

In October 2024, she described startup life as something resembling extreme sport in a career. She also reported strong inbound customer demand. There was a useful tension in those two observations: customers could want the product while the founders were still learning how to deliver it.

Capital, then another kind of test

On October 9, 2025, Datacurve announced a $15 million Series A led by Chemistry. Together with its previously reported $2.7 million seed financing, that brought total funding to $17.7 million. The money was intended to support the provision of training data for further model development.

Ge and Lee had also been recognised in the 2025 Forbes 30 Under 30 AI category. These milestones made the company more visible. They did not settle its underlying product question: how to keep collecting work that teaches a model something useful as the models themselves become more capable.

In 2026, evaluation became a public part of that work. Ge co-authored the introduction to DeepSWE with Wenqi Huang, Lee and Leonard Tng. The benchmark contains 113 original software engineering tasks across 91 repositories and five programming languages. Its purpose is to examine coding agents on longer, more involved work.

The distinction matters. If an agent encounters a solution already present in public material, a score may say less about solving an unfamiliar problem. DeepSWE’s authors describe constructing new tasks and checking observable software behaviour. The benchmark asks for evidence that work has been completed, with attention to how the grading itself operates.

A revised version followed in June. By September, Ge was publicly discussing feedback and further improvements. Measurement, in that account, remained an ongoing project. A leaderboard is easier to print than a reliable definition of what good software engineering requires.

There is a recognisable person running through these different projects: the climber making a training tool, the computer science student adding philosophy, the founder learning a job she had never done. Ge’s declared ambition is to build data infrastructure that helps more capable models emerge. Her earlier writing supplies a useful companion question: what does the work actually do for people?

The breakfast scene gives that question a scale small enough to understand. Thousands of users can fit on a dashboard. One sister handing over one sweet requires a different kind of attention. Ge has written about both. Her company now depends, in part, on understanding the people behind the numbers.