As part of the joint project by Digital Business and Astana Hub, “100 Startup Stories of Central Asia,” zypl.ai Vice President of Product Shukhrat Khalilbekov shared how an AI academy in Dushanbe grew into an international fintech startup and why banks need synthetic data. We also discussed how the company’s algorithms identify reliable borrowers with up to 98% accuracy.
“My Love for Dota 2 Led Me into IT”
— Tell us about yourself: where did you grow up and how did you get into IT?
— I was born in Khujand, a city in northern Tajikistan. At 18, I enrolled at the Russian university Higher School of Economics. I studied economics, but in my second year I took an additional IT specialization. You could say my love for Dota 2 led me there. One day at the university, they presented a new Data Science program and showed visualizations using this game as an example: graphs of relationships between items, transactions between players. It made a strong impression on me. That’s how I started studying data science.
— Where did you work before joining zypl.ai?
— After university, I went to work as an auditor at Ernst & Young. Later, I joined Oliver Wyman, a consulting company that had an internal mini-startup focused on scoring small and medium-sized businesses based on open data. That’s where I first seriously encountered machine learning in finance: we built models that could assess risks and predict a company’s solvency based on its tax identification number.
I changed a few more companies, and then started thinking about launching my own business. Around that time, I met Azizjon Azimi. He once posted a story saying he was looking for a head of product at zypl.ai—someone who understands both business and technology to lead the product and development teams.
I looked at it and thought, “I seem to be a perfect fit.” I wrote to Azizjon, suggested a call, we talked—and that’s how I joined the team.
— What was the company doing at that time?
— At that point, the startup was two years old. Azizjon came up with the idea for zypl.ai while studying at Stanford—he formulated the concept of a platform for assessing credit risks using artificial intelligence. After returning to Tajikistan, he realized there was a lack of specialists capable of implementing such an idea. So he created an AI academy to train engineers and data scientists. The academy still exists today, and the first zypl.ai team emerged from its graduates.
At that time, the focus was on scoring technologies—assessing the creditworthiness of potential bank clients. For example, when a person applies for a loan in an app, the system analyzes the data and decides within seconds whether to approve or reject the application. We were building these technologies—algorithms that process data and deliver decisions.
“We Reduce the Share of Non-Performing Loans by 30%”
— What changed after you joined zypl.ai?
— When I joined the team, the structure was very “startup-like”—without clear processes or defined roles. Even the technology stack varied: everyone worked with whatever tools they preferred. The first thing I did was bring order—structured the team by areas, defined responsibilities, and introduced a product roadmap (a document or visual scheme that outlines what the team plans to do and when — Digital Business note).
This allowed us to move from chaotic decisions to systematic development and focus on key products—primarily zGAN, our synthetic data generation technology, and the Lucid platform, which enables non-technical specialists to build and train ML models.
— Explain in simple terms: why is zGAN needed?
— Imagine a small bank. Every day, employees need to decide who should receive a loan and who should be denied. Doing this manually or using simple rules like “if income is below $1,000—reject” is inaccurate and risky. That’s why banks use machine learning models—algorithms that analyze hundreds of features and predict whether a person will repay a loan.
But for such a model to work well, it needs a lot of high-quality data. Small banks simply don’t have it—there are too few clients and repayment histories. That’s exactly the problem we solve: our tool helps add missing data by generating it synthetically, so the model learns as if the bank had much more information. As a result, it more accurately determines who is likely to repay a loan and who is not.
— And what is the value for large banks? They don’t lack data.
— The point is that synthetic data can be used not only when data is scarce, but also mixed with real data to make the model more robust. zGAN can generate anomalous data, simulating possible but not yet occurred macroeconomic scenarios—for example, sharp inflation spikes or a collapse in commodity markets. We’re talking about so-called “black swans.”
To be clear, we don’t predict specific events like the COVID-19 pandemic. Rather, we create similar situations so that the model understands such deviations are possible and knows how to respond to them.
If tomorrow there is an economic downturn, currency depreciation, or a surge in unemployment, the model will not start making massive scoring errors, because it has already “seen” similar scenarios during training.
In conventional models, accuracy can fluctuate sharply from month to month—95% one month, 70% the next. A model trained with synthetic stress scenarios behaves consistently—and that’s crucial for regulators.
As part of the joint project by Digital Business and Astana Hub, “100 Startup Stories of Central Asia,” zypl.ai Vice President of Product Shukhrat Khalilbekov shared how an AI academy in Dushanbe grew into an international fintech startup and why banks need synthetic data. We also discussed how the company’s algorithms identify reliable borrowers with up to 98% accuracy.
“My Love for Dota 2 Led Me into IT”
— Tell us about yourself: where did you grow up and how did you get into IT?
— I was born in Khujand, a city in northern Tajikistan. At 18, I enrolled at the Russian university Higher School of Economics. I studied economics, but in my second year I took an additional IT specialization. You could say my love for Dota 2 led me there. One day at the university, they presented a new Data Science program and showed visualizations using this game as an example: graphs of relationships between items, transactions between players. It made a strong impression on me. That’s how I started studying data science.
— Where did you work before joining zypl.ai?
— After university, I went to work as an auditor at Ernst & Young. Later, I joined Oliver Wyman, a consulting company that had an internal mini-startup focused on scoring small and medium-sized businesses based on open data. That’s where I first seriously encountered machine learning in finance: we built models that could assess risks and predict a company’s solvency based on its tax identification number.
I changed a few more companies, and then started thinking about launching my own business. Around that time, I met Azizjon Azimi. He once posted a story saying he was looking for a head of product at zypl.ai—someone who understands both business and technology to lead the product and development teams.
I looked at it and thought, “I seem to be a perfect fit.” I wrote to Azizjon, suggested a call, we talked—and that’s how I joined the team.
— What was the company doing at that time?
— At that point, the startup was two years old. Azizjon came up with the idea for zypl.ai while studying at Stanford—he formulated the concept of a platform for assessing credit risks using artificial intelligence. After returning to Tajikistan, he realized there was a lack of specialists capable of implementing such an idea. So he created an AI academy to train engineers and data scientists. The academy still exists today, and the first zypl.ai team emerged from its graduates.
At that time, the focus was on scoring technologies—assessing the creditworthiness of potential bank clients. For example, when a person applies for a loan in an app, the system analyzes the data and decides within seconds whether to approve or reject the application. We were building these technologies—algorithms that process data and deliver decisions.
“We Reduce the Share of Non-Performing Loans by 30%”
— What changed after you joined zypl.ai?
— When I joined the team, the structure was very “startup-like”—without clear processes or defined roles. Even the technology stack varied: everyone worked with whatever tools they preferred. The first thing I did was bring order—structured the team by areas, defined responsibilities, and introduced a product roadmap (a document or visual scheme that outlines what the team plans to do and when — Digital Business note).
This allowed us to move from chaotic decisions to systematic development and focus on key products—primarily zGAN, our synthetic data generation technology, and the Lucid platform, which enables non-technical specialists to build and train ML models.
— Explain in simple terms: why is zGAN needed?
— Imagine a small bank. Every day, employees need to decide who should receive a loan and who should be denied. Doing this manually or using simple rules like “if income is below $1,000—reject” is inaccurate and risky. That’s why banks use machine learning models—algorithms that analyze hundreds of features and predict whether a person will repay a loan.
But for such a model to work well, it needs a lot of high-quality data. Small banks simply don’t have it—there are too few clients and repayment histories. That’s exactly the problem we solve: our tool helps add missing data by generating it synthetically, so the model learns as if the bank had much more information. As a result, it more accurately determines who is likely to repay a loan and who is not.
— And what is the value for large banks? They don’t lack data.
— The point is that synthetic data can be used not only when data is scarce, but also mixed with real data to make the model more robust. zGAN can generate anomalous data, simulating possible but not yet occurred macroeconomic scenarios—for example, sharp inflation spikes or a collapse in commodity markets. We’re talking about so-called “black swans.”
To be clear, we don’t predict specific events like the COVID-19 pandemic. Rather, we create similar situations so that the model understands such deviations are possible and knows how to respond to them.
If tomorrow there is an economic downturn, currency depreciation, or a surge in unemployment, the model will not start making massive scoring errors, because it has already “seen” similar scenarios during training.
In conventional models, accuracy can fluctuate sharply from month to month—95% one month, 70% the next. A model trained with synthetic stress scenarios behaves consistently—and that’s crucial for regulators.
Share
