"I do not see problems to be a female founder." The secrets of the most successful AI startup with Russian roots

This material (information) was produced and distributed by foreign agent Elizaveta Nikolaevna Ossetian or concerns the activities of the foreign agent Elizabeth Nikolaevna Ossetian 18+

From childhood, Anna Dorogush loved mathematics, graduated from the faculty of computational mathematics and cybernetics of Moscow State University and Yandex School, after which she left for the USA to work in Microsoft. Then she returned to Moscow and worked first in the Moscow office of Google, and then in Yandex, where she created one of the most popular Catboost libraries. In March 2022, Dorogush resigned from Yandex and launched her own startup, which is already called the most successful AI startup with Russian roots. To date, RecRAFT has attracted a total of $ 50 million from Khosla Ventures, Accel, RTP and other large funds, and the model itself is one of the top three neural networks for generating images. It is noteworthy that at the time of launch and departure from Russia, Anna had a three -year -old child, and right now she is waiting for the second.

Elizabeth Ossetian (recognized as an ino -agent) came to London to find out how to come up with an idea for a successful startup, where to attract investments without a team and a finished product, which has changed in the field of design with the advent of AI and how to teach children to love mathematics. And also - about motherhood and business, attitude of investors to pregnancy, the work of neural networks and 5 million users without advertising.

We give short excerpts from the interview. Watch it entirely on the channel "This is Ossetian."

Career path

- You studied at the German humanitarian school, and mathematics became your specialty. How did it happen?

- My parents are mathematicians. Mom worked as an assistant professor at the Moscow State Polygraphic University. Dad, unfortunately, died when I was small. He was a very good mathematician. I probably just liked mathematics since childhood. No one specifically pushed me to engage in mathematics, and, unfortunately, I did not participate in the Olympics - I simply did not know about them.

I studied at a German school with a humanitarian bias and in -depth study of the German language, but I always liked mathematics. Since childhood, I knew for sure that I want to deal with mathematics.

- Why did you choose the Navy [Faculty of Computational Mathematics and Cybernetics. - approx. The Bell] Moscow State University?

- I chose between VMK and Mehmat. Mom said that if mathematics, it means a mechmat, and I decided to go in my path and chose the VMK. This was a kind of protest.

Parents had no resources to pay for their studies, but it was possible to take a loan. The first course I studied perfectly - only five. From the second year she began to work as a model to pay for training. The work, of course, interfered with my studies, and now I think it was a big mistake.

“Do you think this greatly damaged your studies?”

- I always loved to study and now I love very much. If there is an opportunity in a student to receive knowledge, go to lectures and solve problems, you need to use it. Passing is very disappointing. Then there is no such opportunity.

After Moscow State University, I entered the Shad [Yandex School of Analysis School. - approx. The Bell]. I believe that I owes my whole career first of all, this is one of the best educational institutions in the world in the field of machine learning. There is a very strong teaching staff, complex and interesting tasks.

-At some point, did you leave to work in Microsoft in Seattle?

- Yes, but there are several more small cities nearby: Redmond, Kerkland and Belve. In these cities there are offices of large companies. Therefore, many employees do not live in Seattle, but in one of these cities. I lived in Kerkland.

And Seattle, and Belveu, and Kerkland are beautiful cities, there is cool there. And in Microsoft I also really liked working. But I decided that I would like to return to Russia for personal reasons.

- Do you regret that you haven't stayed in America?

- No, I do not regret. When I moved to America, I had several offers: I could go to the States or to London, to Google. Then I chose America. Since then, Google was regularly called in Google - they constantly maintain contact with candidates. I worked in the Moscow office of Google for about a year until it was closed. After that, I moved to Yandex.

In Google and Microsoft, I worked as a linear developer. In Yandex, it was also a linear developer, but there is my main career growth. Most of the things that I have learned are in Yandex.

- And how do the developer get to work in Microsoft?

- When I got in Microsoft, they conducted the so -called Hiring Trip: several Microsoft teams came to different countries and conducted a lot of interviews in a row in one or two weeks. Those who have passed a successful interview are invited to work.

Now, after the pandemic, the interviews for the most part have moved online. Therefore, I am not sure that Hiring Tripa is just as popular now. But then it was an effective way.

- Why is Google so actively Huntil?

- I think this is just part of the active work of their recruiters. They write to many graduates of top universities. Then Shad was not yet widely known, but Moscow State University - yes.

- What did you do in Yandex?

- At first I was engaged in various tasks related to ranking and search. Then there were tasks about the matrixnet on the gradient boosting. And then I already began to tightly engage in gradient boosting.

- Matrix ... what? Gradient boosting?

- This is an algorithm for machine learning, which works well on tabular data. Now almost everything is solved by neural networks, but the gradient boost is still - a top -end algorithm for solving a large number of problems on tabular data.

-And if you explain in a very simple way? Is this a big formula? Or a big program?

- We can say that this is a large training formula. She, like a neural network, accepts data to the input, and gives a number to the output - prediction. For example, in the search, this may be a prediction of the ranking of the page: if one page is assigned a value of 5 and the other - 4, then the first will be displayed higher in the search results. You can predict the cost of housing. But, like any algorithm, there are mistakes, although the gradient boosting of the quality metric in different modes is higher. And just use it.

- Have you worked in Yandex for about six years?

- Seven. I advanced like a leader: I developed a gradient boosting, added a lot of new things here, only my efforts were not enough, and I created a team. Gradually, we began to engage in other libraries of machine learning (ML). As a result, the area of ​​responsibility was formed: libraries and tools for ML. Other teams and ML engineers came to us to make changes or use our solutions. This is probably my main career path and success in Yandex.

- Why did you decide to leave?

“I thought about starting my business for quite some time.” During my career, I launched several successful products from scratch - both in Yandex and Google. But all this was inside the companies. And I wanted to build something of my own from scratch. I have always perceived myself as a grocery person. Now I understand how much I did not know much - and, perhaps, I still do not know. But at that time I was sure that I was ready.

- How did you report this?

- Just left. There were no questions. I quit in March 2022. Even earlier, in 2021, she planned to move to the United States, prepared an O-1 visa-a “talent visa”. I have a good track record, an understandable “talent sheet”. I received O-1 and could move to the USA alone. But to transport the team was a very difficult task. Therefore, I chose Dubai. We moved the team there, lived there a little less than a year, and then moved to London.

Launch RecRAFT

- You were engaged in things far from B2C and design as a whole. But suddenly you leave Yandex and begin to make a company ... in fact, perpendicular to everything that you did before, although based on AI.

- I made an internal decision that I want to try to build my product outside a large company. I thought about it for many years, and still took a step. At first I just tried to understand for several months, but what exactly to do? This is a fairly frequent template among the faunters. It is believed that successful companies are born from a deep knowledge of the subject. But in most cases - a person first decides to build a company, and then begins to explore.

I was going to make a product in the Machine Learning Tools [software that allows you to create and train AI models for data analysis. - approx. The Bell]. This is what I knew. I began to communicate with potential users, to explore the market.

Around that moment, I was engaged in Rust, AI wave began. Midjourney launched, articles on the diffusion generation of images appeared, and it became clear that in the world of design, art and everything that concerns images, and in the future and video, everything would change very quickly. And this is the moment when you can participate in the formation of future design and new professions.

- How to understand that now is that unique moment? It often seems that everything has already been invented.

- Now - this is the very moment. Still. All industries with the advent of a new technology are changing greatly, and then how they will look, it depends very much on what tools we will build.

My younger sister is a designer. Therefore, it was easier for me to start understanding this area. I just called up to Zoom with her, she rummaged the screen, showed how she was working. And I watched.

I remember how I noted that my sister spends a lot of time on vectorization. This is the transformation of a raster picture - pixels - into a vector form, where the picture consists of shaps (circles, rectangles). And I had the first thought: "How many are the routers of the routine, let's help to automate it."

But soon I realized: designers love their routine very much. The process itself is important to them, craft.

- Love their craft - from this [the name of the startup] recraft?

-Yes, recraft, of course, is associated with the word “craft”, although I have not chosen this word the first time.

I had a lot of conversations with my sister and other designers. I wrote to them in LinkedIn, asked questions on sites where you can communicate with experts, asked for an interview. This helped me understand how, in principle, this area is arranged.

Then I decided that we will adapt the technology of images for professional scenarios. The creation of Vorkflow is a place where it is most convenient for designers to work. RecRAFT is an endless Canvas, on which you can have many layers, make a story, compare pictures. This is the working environment familiar to designers.

“Maybe I will say muck, but this is something like Photoshop?”

- In terms of Workflow, the endless Canvas is a thing that looks more like Figma, but not about grocery, but graphic design.

From the very beginning, we had three key priorities: quality, convenient for Workflow and functionality, specific to design. We started with a vector-art and decided as follows: "No one knows how to generate it, and we will become the first to learn." And steel. This was the first version of the product, we positioned something like this that we are making a product for designers in which designers can generate a vector.

The vector is important in many scenarios, but this is far from all that is needed. Why did we first choose such positioning? Because it was very different from everything else.

- What already happened?

- From everything else, from generating just images. But this played a cruel joke with us. For a long time, people thought that RecRAFT is exclusively about vector art, and that’s all. Although it was only a small fragment of everything that we proposed.

-The market is large?

- Only in the USA - hundreds of thousands of professional designers. Their number around the world is growing, because the design, including with the advent of AI, becomes in some cases a more affordable profession.

- That is, AI reduces the entrance threshold?

- This is a difficult question. I think that we are now at the very early stages of technology development.

- Recraft earns?

- Now - a little more than $ 5 million per year. We turned on the Credit-Based subscription in September last year and have been growing quite quickly since then.

- That is, $ 5 million is revenue in about 9 months?

- Yes. We have two parts of business. The first is subscription. These are mainly separate designers or companies that make a subscription for a month or for a year. The second is the API. It is used by companies when you need to generate images in large volumes (in Adtech, Marketing, SEO).

- How many users do you have?

- Close to 5 million registered users, active per month - less. In AI, in general, there is a phenomenon that we call AI-Turism-when people just come to try. There are many such users too.

- As I understand it, the earnings are far from covering Burn Rate [income does not cover all expenses]?

- For today - yes. There are several large articles of expenses: training training, the use of models (Inference), salaries and marketing.

Advertising has not yet paid off. We really did not optimize monetization and are now more focused on building a product.