Vahan Petrosyan and the trouble with one good algorithm

Vahan Petrosyan had an algorithm, a prototype and evidence that companies wanted what he had made. Some offered to buy it. At a computer-vision conference in 2018, a piece of doctoral research began to look like a business. He called his brother Tigran. Soon, both would leave their PhD programs. The difficult part of the experiment was about to begin.

The algorithm addressed a precise frustration: teaching a computer what belongs inside the boundary of an object. Before a machine can learn from a picture, somebody often has to identify the things in it. A car. A road. The pixels around an awkward edge. Petrosyan’s research at Stockholm’s KTH Royal Institute of Technology offered a way to accelerate that work. It was an attractive beginning for SuperAnnotate, the company he would co-found.

Years later, he would describe the original belief that one algorithm could sustain a business as mistaken. The admission gives his career its shape. Petrosyan became a founder because the research worked. He became responsible for a growing company by learning how much else had to work around it.

A conference changes the assignment

Initially, he had imagined using his segmentation technology for image-editing applications. A person choosing part of a photograph is, after all, dealing with a related problem: finding the right region without laboriously tracing every edge. At CVPR in 2018, he saw another market in the companies preparing training data for machine learning.

Here was work whose value increased when other people could do it accurately and quickly. His prototype attracted attention from data-labeling companies. Selling the underlying algorithm was one possibility. In his later account, the response prompted a different decision: build a company with his brother. The research question had acquired a customer.

There is a concrete record of the technical work behind that moment. A US patent titled “Superpixel merging”, granted in March 2023, names Vahan Petrosyan as its inventor. The method concerns combining small regions of an image into larger ones using information about similarity and boundaries. It is a rather literal foundation for a business that would spend years helping people organize information.

The cheque before the venture rounds

In 2018, the brothers took their idea to Armenia’s Sevan Startup Summit and won. A photograph from the event shows Vahan on the left and Tigran on the right, holding a large board announcing a $5,000 prize. For once, a startup’s early promise came with handles, or at least enough cardboard to require two people.

The brothers returned to Yerevan to build the startup. Granatus Ventures became an early backer. Tigran later explained that early investment helped them attend conferences, make contacts and give the young company a presence. Getting noticed required showing up, which required money before there was much company to notice.

Vahan initially led technology; Tigran was chief executive. Their shared decision to leave academic research was followed by a division of responsibilities. An invention had become something that needed a product, customers and an organization. The family connection supplied a co-founder. It did not remove the unfamiliar work.

Stockholm supplies more than an algorithm

Petrosyan had contacted KTH Innovation in 2017. With Tigran and KTH student Oscar Örnberg, he developed the tool and worked through the university’s startup program. Its coaching helped the team examine applications beyond the first image-editing idea. Workshops and conversations supplied a different vocabulary from the one needed to explain a research result.

When Petrosyan looked back in April 2025, he singled out help with the business model, patenting and initial investment from KTH Holding. Those details complicate the familiar image of a researcher suddenly deciding to become an entrepreneur. There was a support system around the decision, with people helping work out what might be commercially useful.

The company’s path would cross several borders, but the Swedish connection remained part of its story. So did Örnberg’s early collaboration. Petrosyan’s algorithm mattered; the people who helped turn it into something usable deserve space in the picture too.

Berkeley introduces the customer

In January 2019, the company joined Berkeley’s SkyDeck accelerator and moved its headquarters to Silicon Valley. Vahan’s account of the accelerator period includes an introduction to an autonomous-driving company that became an early major contract. His image technology fitted a market then investing heavily in computer vision. A prospective application had become a paying customer’s task.

That brought a different sort of validation. A demonstration could show that an algorithm was fast. A contract meant someone had a problem worth paying to solve. Petrosyan has described the period as a change in mindset, from continuing research to learning how to grow a company.

“We were research people who had to turn into business people.”

Vahan Petrosyan · 2025 interview

SuperAnnotate publicly released its image-annotation platform in April 2020. A $3 million seed round followed; a $14.5 million Series A arrived in 2021. The sequence gave the original research room to develop into software people could use together. Annotation involved assigning work, checking quality and preserving what had been produced, as well as selecting regions in pictures.

In August 2020, Vahan and Tigran appeared together on Software Engineering Daily. The discussion covered an annotation platform that other annotation businesses could use. Even then, the product occupied an interesting position: it could supply the tools for companies doing the labeling, alongside serving teams that needed the resulting datasets. The customer could itself be in the annotation business.

By 2022, Petrosyan was describing ground-truth data as something machine-learning engineers needed to manage with the care software developers gave their code. Creating labels was one task; keeping track of the data and its versions was another. His comparison with GitHub made the ambition easier to understand. The company wanted a recurring place in the engineer’s work, with the accumulated dataset requiring attention after an initial round of labeling.

In his 2025 discussion of the business, Petrosyan emphasized another change: making it possible for enterprises to work with multiple service providers through one platform. That required seeing providers as potential collaborators. His ambitions had expanded to the infrastructure that could connect their work with internal customer teams.

He also gave solution engineers a central place in selling enterprise AI software. Sophisticated tools need people who can help customers put them to use. It is a practical concern for somebody whose starting point was a research prototype: usefulness has to survive the journey into another organization.

The job grows around the research

Petrosyan became CEO in 2023, after serving as CTO. That August, SuperAnnotate launched software for generative AI. The company was working on a broader set of data problems, encompassing text, images, audio and video, as customers began building with large language models.

In his November 2024 funding announcement, he described enterprises trying to adapt older tools, build their own systems or handle complex projects in spreadsheets. The opportunity he identified was to help turn proprietary information into usable training data, with flexible workflows for different applications. A public benchmark could tell a customer something about a model; testing on the customer’s own material could answer a more specific question.

One connection led back to Sweden: Databricks co-founder Ali Ghodsi had previously been an assistant professor at KTH. By 2025, Databricks was among SuperAnnotate’s customers. The university that helped Petrosyan begin had another link to a company now using what his team had built.

Dell Technologies Capital subsequently invested $13.5 million, extending the Series B to approximately $50 million. In June 2025, SuperAnnotate was also recognized as a Databricks customer-impact partner. The figures describe a company financing round. The more revealing continuity in Petrosyan’s career is his attention to the work needed before an AI system can be judged useful.

He has also offered an unexpectedly relaxed observation about fundraising. “Whenever I was actively fundraising, I was doing something wrong.” In that 2025 conversation, he described successful investor discussions beginning with informal coffees. It is his account of his experience, with a hint of the distance between how fundraising is planned and how a relationship actually begins.

A statistician in the room

For a personal introduction to Petrosyan, try his reading recommendation: The Elements of Statistical Learning. Asked about his favorite mathematical paradox in 2022, he chose Simpson’s paradox. His earlier interests included financial and actuarial mathematics, quantitative economics, statistics and data visualization. He traced his introduction to machine learning to a course taught by Adele Cutler, a co-creator of Random Forests.

These details give the executive title some texture. The researcher did not disappear when the job changed. His interviews still return to the mechanics of data quality: organizing examples, evaluating results and repeating the process as requirements change.

By 2025, he was discussing the difficulty of judging AI answers when several responses might be acceptable. Customer preferences can shift; a model’s behavior may need to shift with them. In his description of data operations, evaluation and monitoring belong to an ongoing cycle. The boundary that matters may now be a company’s definition of a satisfactory answer.

In February 2026, the Armenian Innovation Foundation appointed him to its board. The connection to Armenia had been present since the early competition, the return to Yerevan and the investors who backed the brothers. It now included a formal role beyond SuperAnnotate.

The photograph from Sevan still offers a pleasingly concrete place to end: two brothers, a prize board and a business whose next problems had yet to announce themselves. Petrosyan had found a way to make part of a picture easier to select. Building the company meant keeping more of the picture in view.