The startup zypl.ai was founded in 2021 by Tajik entrepreneur Azizjon Azimi. The company develops technologies that help banks and microfinance institutions reduce credit risk. Their model can predict who is likely to repay a loan and who shouldn’t be approved. Today, zypl.ai operates in 20 countries, including Kazakhstan, and partners with 60 financial institutions. The latest confirmed valuation of the company is $40 million.
As part of the joint project by Digital Business and Astana Hub, “100 Startup Stories of Central Asia,” Shukhrat Khalilbekov, Vice President of Product at zypl.ai, shared how an AI academy in Dushanbe evolved into an international fintech startup and why synthetic data is essential for banks. We also talked about how the company’s algorithms can identify reliable borrowers with up to 98% accuracy.
«I got into IT thanks to my love for Dota 2»
— 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 Higher School of Economics in Russia, majoring in economics. But in my second year, I picked up an additional focus in IT. You could say it was my love for Dota 2 that led me there. One day, the university presented a new Data Science program and used the game as an example, visualizing things like item networks and player transactions. That really stuck with me. That’s when I started diving into data science.
— Where did you work before joining zypl.ai?
— After university, I started out as an auditor at Ernst & Young. Later, I joined Oliver Wyman, a consulting firm that had an internal mini-startup focused on scoring small and medium-sized businesses using open data. That’s where I first got hands-on experience with machine learning in finance. We built models that could assess risk and predict a company’s creditworthiness just from its tax ID number.
I worked at a few more companies after that, and eventually started thinking about launching something of my own. Around that time, I came across Azizjon Azimi. He had posted a story saying he was looking for a Head of Product at zypl.ai, someone who understands both business and tech, and could lead the product and development teams.
I saw the post and thought, “This sounds like a perfect fit for me.” I messaged Azizjon, suggested we jump on a call, and we talked. That’s how I joined the team.
— What was the company working on at that time?
— At that time, the startup was two years old. The idea for zypl.ai came to Azizjon while he was studying at Stanford. That’s when he developed the concept of a platform that uses artificial intelligence to assess credit risk. When he returned to Tajikistan, he realized there weren’t enough specialists in the country to bring that idea to life. So he launched an AI academy to train engineers and data scientists. The academy still exists today, and its graduates became the first team behind zypl.ai.
Back then, the focus was on credit scoring technology, assessing the creditworthiness of potential bank clients. For example, when someone applies for a loan through a mobile app, the system analyzes their data and decides within seconds whether to approve or reject the application. We were building that tech, developing algorithms that process the data and deliver a decision.
«We reduce the share of unpaid loans by 30 percent»
— What changed after you joined zypl.ai?
— When I joined the team, the structure was very much in startup mode. There were no clear processes or role distributions. Even the tech stack varied from person to person — everyone used whatever they were most comfortable with. The first thing I did was bring some order. I organized the team by function, defined areas of responsibility, and introduced a product roadmap (a document or visual plan that lays out what the team is going to build and when — note by Digital Business).
That helped us shift from scattered, ad hoc solutions to a more systematic development process. We were able to focus on our key products, starting with zGAN, our synthetic data generation technology, and Lucid, a platform that lets non-technical users build and train machine learning models.
— Can you explain in simple terms what zGAN is for?
— Imagine a small bank. Every day, employees have to decide who gets approved for a loan and who gets denied. Doing this manually or using simple rules like “if income is under a thousand dollars, then reject” is risky and not very accurate. That’s why banks rely on machine learning models. These are algorithms that can analyze hundreds of factors and predict whether a person is likely to repay the loan.
But for a model like that to work well, it needs a lot of high-quality data. Small banks usually don’t have enough of it, because they don’t have many clients or enough repayment history. That’s exactly the problem we solve. Our tool helps fill in the gaps by generating synthetic data, so the model can train as if the bank had much more information. As a result, it becomes more accurate at predicting who is likely to repay a loan and who is not.
— What about big banks? They don’t seem to have a shortage of data, so where’s the value for them?
— The point is, synthetic data isn’t just useful when you’re short on real data. You can mix it with actual data to make your model more resilient. zGAN can generate rare or extreme cases, simulating macroeconomic scenarios that haven’t happened yet, like sudden spikes in inflation or a collapse in commodity markets. This helps banks prepare for so-called “black swan” events.
Let me clarify right away: we don’t try to predict specific events like the COVID-19 pandemic. It’s more about simulating similar scenarios so the model learns that such deviations can happen and knows how to respond when they do.
If there’s an economic downturn tomorrow, a currency crash, or a spike in unemployment, the model won’t start making a flood of mistakes in credit scoring because it has already “seen” similar scenarios during training.
In regular models, accuracy can fluctuate from month to month — one time it’s 95%, then it drops to 70%. A model trained with synthetic stress scenarios behaves more consistently, and that kind of stability is crucial for regulators.
— Your second major product is Lucid. What problem does it solve?
— A bank or microfinance employee can upload their data to the platform and choose a task, like credit scoring or fraud detection. Lucid then builds the model, checks its metrics, and visualizes the results. After that, the user can test it on real data, compare it with their current performance, and if needed, integrate the new model into their business processes.
Basically, it’s a no-code tool that makes it possible for smaller organizations to use AI, not just large companies with in-house ML teams. It opens the door for businesses that don’t have those kinds of resources.
— How do you measure the effectiveness of your products?
— We focus on key metrics. For credit scoring, it’s the non-performing loan rate (the share of bad loans). If that number goes down, it means the model is working. For fraud detection, it’s about the number of fraudulent cases caught and reducing false positives.
On average, we reduce the share of unpaid loans by 30%, which brings a noticeable economic benefit. When a bank issues a loan, it’s required to keep part of that amount in reserve as a safety cushion in case of defaults. But those frozen funds represent lost profit.
Typically, the drop-off rate, which is the difference between the model’s prediction and the actual result, is under 2%, and usually around 1.5%. In some regions, it performs even better.
«The first reaction was: AI? Scoring? Please leave»
— Let’s talk about your clients. How many do you have, and in which countries are they located?
— We’re active in more than 20 countries. Our reach is pretty broad, covering the Middle East, Africa, Southeast Asia, the US, and Latin America. In Kazakhstan, for example, we already have several microfinance partners and are finalizing a contract with a major bank.
We work with around 60 organizations, including banks and microfinance institutions. Using zypl.ai’s technology, banks around the world have already issued over $400 million in loans.
We recently had an interesting case in a new industry for us — commodity trading. The clients were companies that deal in trading raw materials like grain, oil, and metals. To put it simply, they don’t buy the physical goods but rather securities backed by these assets. For them, we developed models that help assess deal risks using both historical and synthetic data.
— How do you find clients, considering that banks are traditionally cautious about adopting new tech?
— The banking sector is very risk-averse, so the first two years were the toughest. When the team first started talking to banks in Tajikistan about credit scoring and machine learning, the reaction was something like, “AI? Scoring? What are you even talking about? Please leave.”
Our first two clients only agreed to work with us after months of negotiations. But once they saw real results six months later, including growth in their loan portfolios and reduced risk, the industry’s attitude started to shift. That’s when the product began spreading actively across the region.
After those early wins, we developed a clear and transparent approach for onboarding new clients. We don’t try to convince them with presentations. Instead, we invite them to test everything in practice. We say, “Don’t take our word for it. Let’s run a model test on your historical data. If your current default rate is 7%, we guarantee our algorithm can reduce it by 30%.” The bank sends us old, already processed loan applications — cases where the outcomes are known. We run that data through our model and then compare the predictions to the actual results. After the test, they see how much more accurate the forecasts are, and then decide whether to work with us.
— How does zypl.ai make money?
— We have three main monetization models.
The first model is pay-per-request. When a loan application comes into the system, we process it and charge a fixed fee for each request. The price depends on the size of the portfolio. The more applications there are, the lower the cost per request.
The second model is a fixed subscription. A financial organization pays a set amount monthly or annually, with no limits on the number of requests. The pricing also depends on the size of the client. We offer discounts to smaller companies to ease the pressure on their budgets.
The third model is success fee-based, and it’s a new approach we introduced this year. The client only covers operational costs upfront. After a set period, usually every six months, we review the results together. If the agreed-upon KPIs are met, we receive a percentage of the savings generated.
«We’re very close to breaking even and plan to reach that point by the end of this year»
— Is the company already profitable or still on the path to profitability?
— We have a strong ARR, meaning annual recurring revenue. We’re very close to breaking even and expect to reach that point by the end of this year or in the first quarter of next year. We might even start turning a profit.
— How much money has been invested in the project, both personal and venture capital?
— The company has always grown through external funding. The total amount raised is around $9.4 million, including the most recent round of $6.3 million at a $40 million valuation. The lead investor in that round was the European holding company Prosus Ventures.
— What was the purpose of the funding round in 2025?
— At the time of the deal with Prosus Ventures, we already had enough funds in our accounts to operate comfortably for several more years without raising new investments. We also had stable revenue. But we realized that to reach the next level, we needed more than just money. We needed reputation and connections. When you approach potential partners and mention that Prosus Ventures is one of your investors, the reaction is completely different. People immediately see that this isn’t just a random startup but a serious company.
Still, the funding came in handy for two main goals. The first was expansion. That requires a strong sales and partnerships team. Until recently, our sales were handled by top management and the product team. Leadership was shared between Azizjon, myself, Mihir Modi (our VP of Strategy), and our COO. Each of us oversaw several areas. Azizjon and Mihir focused on international and large-scale deals, while I handled the product side across all regions with the product team. Now we’ve hired our first VP of Business Development, a person with extensive experience working in credit bureaus around the world. We’re currently building a full international sales team around him.
The second goal is R&D. We have several technological ideas that require a very high level of expertise. And we can’t implement them using only local specialists. Engineers with the necessary skills are rare and come at a high cost.
— What are the company’s upcoming plans?
— To scale and keep developing the product. We plan to introduce new AI models and expand into related areas, such as guaranteed lending.
In the short term, we’re focused on attracting investment. We plan to close our Series A round in the first half of 2026. In the long term, we’re considering going public. For now, our main focus is on growth, launching new products, and strengthening our position in key markets.