Yandex
ML Prize

As enrollment grows, she continues to develop the program itself: refreshing the content, building flexible learning paths, and integrating AI tools into the learning process.

The program covers every stage in the life of an ML solution: collecting and processing data, training models, deploying them to high-traffic services, and maintaining them with MLOps. This training allows graduates to choose different career paths, such as ML engineering, data science, or research.

Under his direction, the program has grown beyond classical robotics: mechanics and control now sit alongside machine learning and computer vision.

Sergey also founded an international robotics lab where students continue their research and projects.

The course covers neural ranking, transformers, graph neural networks, and generative models. Much of it is hands‑on research: students dig into recent papers, reproduce the methods described in them, and learn to form and test their own hypotheses.

It runs for four semesters and pairs MIPT’s mathematical grounding with today’s industry challenges. Students study statistics, machine learning, deep learning, and applied analytics. They work through the mathematics behind machine learning methods, explore neural network architectures, and learn to check that their conclusions hold up in practice.

The course keeps pace with generative AI: students study diffusion, multimodal, and 3D models, and their homework puts state-of-the-art techniques from the last few years’ research into practice.

In two years, VisualGenAI grew from a one-week intensive into a full semester-long course. Its audience reaches well beyond Yandex School of Data Analysis students: almost as many people now enroll in it on their own as come from the School.

His students derive each algorithm from the math by hand, probe its limitations, and build it from scratch. That grounding lets them pick up new methods on their own, even ones only just surfacing in research papers and not yet in any textbook.

Konstantin has also written a book on reinforcement learning and its use in large language models, now being translated into Chinese and French.

There are almost no traditional lectures. Instead, students tackle 12 assignments individually, then come together to compare approaches and results in class.

In a single course, students work with tabular, text, and geospatial data while learning statistical methods and classical machine learning models. Presenting results clearly and explaining conclusions gets special attention.

Taisiya built the course from her own experience and from interviews with practicing analysts to bring the learning experience closer to real data work.

More than 3000 learners enrolled.

That experience grew into the Design of Modern LLMs course at HSE University. It takes students from the basics of large language model architecture all the way to building their own solutions: preparing data, fine-tuning models for specific tasks, connecting external data sources, evaluating quality and safety, and carrying their projects through to working prototypes.