In this interview, Andrey Krotkikh, Senior Machine Learning Engineer at AliExpress, provides valuable insights into the role of AI in e-commerce. He discusses the design of dynamic pricing systems, the evolution of cashback programs through machine learning, and the growing importance of automation in marketplace operations. Andrey also delves into how AI models are improving customer support forecasting and collaboration across teams while offering his perspective on future AI trends that will shape the industry. Read on for a closer look at how these innovations are driving efficiency and growth at AliExpress.
What are the key considerations when designing dynamic pricing systems for a global marketplace, and how do you mitigate risks such as customer dissatisfaction or market fluctuations?
Dynamic pricing models must be sensitive to external events like holidays, global promotions, and seasonal trends that impact customer behavior. To do this effectively, the model relies on gathering wide-ranging data over a long period to capture patterns in demand and customer responses to price changes. This includes analyzing historical trends, seasonal shifts, and responses to specific triggers.
By using time-series analysis and factoring in external signals, the model can better predict how events influence buying behavior. For example, demand may spike during major promotions, but price elasticity can shift, requiring the model to adapt. To keep things stable, it’s important to have safeguards in place, like limits on sudden price changes, to avoid frustrating customers.
How have cashback and loyalty systems evolved with the integration of machine learning, and what metrics do you use to measure their effectiveness?
With machine learning, cashback and loyalty systems have become more controllable and goal-oriented. Instead of offering generic rewards, these programs now analyze customer behavior and transaction data to create incentives tailored to specific objectives, like boosting GMV, improving retention, or increasing conversion rates. Machine learning identifies the rewards that deliver the greatest impact — whether reactivating inactive users or encouraging repeat purchases — while carefully managing costs and outcomes.
To measure success, key metrics include Customer Acquisition Cost (CAC) and Return on Investment (ROI) to ensure rewards remain cost-effective. Other metrics like incremental GMV and customer lifetime value (CLV) help assess the long-term impact of these programs.
What role does automation play in optimizing marketplace operations, and where do you see the greatest potential for further AI-driven efficiencies?
Automation is irreplaceable for streamlining marketplace operations. It makes testing, prototyping, and deploying machine learning models easier and faster. It also allows teams to take insights derived from data and experiments, quickly implement them into production, and validate hypotheses in a matter of days or weeks. For example, with a well-automated pipeline, a new pricing model or recommendation strategy can be deployed within 1-2 days, tested over 1-2 weeks, and results can be analyzed immediately. This speed of iteration is impossible to achieve manually.
In forecasting customer support issues, what machine learning models or approaches have proven most effective, and how do they enhance the overall customer experience?
Statistical models are highly efficient in forecasting customer support issues. This, now, is a well-known fact: they are simple and easy to deploy. They can quickly spot patterns in historical data, like surges in customer complaints or support queries, without needing complex infrastructure or extensive training. Their transparency makes predictions easy to understand and validate, making them practical for real-world use.
Companies can proactively manage customer service by accurately forecasting when and where issues might arise, reaching faster response times and higher customer satisfaction. Those satisfied customers are less likely to escalate issues or file disputes, which helps cut operational costs related to refunds, returns, and lengthy support tickets.
How do you balance the trade-offs between innovation and scalability when implementing new AI solutions across a platform as large as AliExpress?
I prioritize starting with simple baseline models because they’re easier to explain to business stakeholders. This helps them understand the logic, monitor processes, and align AI solutions with business goals.
Innovation is applied in target areas, like using AI to accelerate development, enabling faster deployment of new features and supporting platform growth. This balance ensures scalability while also improving efficiency and adaptability.
What are the most significant challenges you’ve encountered in deploying AI models in real-world marketplace environments, and how have you addressed them?
One major challenge I faced was deploying a cutting-edge model like TFT (Temporal Fusion Transformer) for time-series forecasting. The model’s complexity made it difficult to manage due to infrastructure and deployment limitations. The experience highlighted the gap between advanced models and real-world constraints, such as computational resources, maintenance, and scalability.
After that, I shifted to integrating simpler, explainable models with straightforward logic, making them easier to integrate, align with market needs, and provide clear, trusted insights. This approach enables faster deployment, easier monitoring, and reliable performance without straining the infrastructure.
What trends or breakthroughs in machine learning do you believe will most influence the future of e-commerce and marketplace operations in the next five years?
I believe large language models (LLMs) will significantly impact e-commerce and marketplace operations in the coming years, as their integration into business processes is still in its early stages. Early applications will likely focus on straightforward tasks like automating customer interactions through chatbots or enhancing search and recommendation systems. Over time, more advanced uses will emerge as businesses explore the full potential of LLMs.
The impact won’t stop at customer-facing tools. LLMs will also boost operational efficiency across various processes. For example, they can automate content generation for product descriptions, streamline customer support workflows, and improve demand forecasting by analyzing unstructured data like reviews and inquiries. As these models advance, they’ll help businesses enhance customer experiences and internal operations, driving efficiency and scalable growth.
As a Senior Machine Learning Engineer, how do you ensure cross-functional collaboration between engineering, product, and business teams to align AI innovations with marketplace goals?
As a Senior Machine Learning Engineer, I prioritize cross-functional collaboration by deeply understanding the end-to-end process that delivers value to customers. I identify bottlenecks and potential issues that could hinder progress and leverage a “T-shaped” skill set — broad knowledge of business processes, data engineering, and analytics, paired with deep expertise in model quality, deployment, and prototyping.
This approach enables me to communicate effectively with engineering, product, and business teams, aligning AI innovations with marketplace goals. By bridging these teams, I ensure that AI solutions are practical, scalable, and contribute directly to the company’s success.