Building AI that starts with research and ends in real world impact.

I am a high school student and AI researcher from Aktobe, Kazakhstan. I started working on AI because I wanted to understand how the models I was using actually worked.

That curiosity led to KazBERT. I adapted BERT for Kazakh with a custom tokenizer and continued masked language model training, then released the model publicly. It began as a learning project. Later, I found researchers using it in their own work, which I had not expected when I started.

In 2026, Agzam Shamsadinov and I entered SemEval Task 13, Subtask C, a shared task on detecting Human, AI generated, Hybrid, and Adversarial code. I designed the Dynamic Multiple Instance Learning pipeline for long code files and worked on EDA and handcrafted features. Our system placed first with 0.7855 macro F1, and our paper was published in the SemEval proceedings at ACL 2026.

Research is difficult to enter when you are still in school, especially outside a major research center. I started DSML Young so that students like us could share experiments, get honest feedback, and meet mentors who take their work seriously. The community now includes students from Kazakhstan and Russia.

The problems I find most interesting are usually the ones that do not fit a clean benchmark. A code file is longer than the model can read. A language has little training data. An agent has to make a sequence of decisions instead of producing one answer. My current interests in coding agents, AI generated content detection, and multilingual AI grew from those cases.

I have also worked on robotics and edge AI. Working with real hardware made me pay attention to things that model evaluations often ignore, especially limited compute and imperfect sensor data. I want to build systems that remain useful when conditions are less controlled.

Selected Projects

A few projects I keep coming back to:

• SemEval-2026 (Task 13) - 1st place system for Subtask C; my work focused on MIL, EDA, and feature engineering.
• Sentra Retail OS - edge-first retail analytics platform for camera-based spatial intelligence.
• KazBERT — a BERT model pretrained from scratch with a custom tokenizer for Kazakh. Cited in MDPI and Frontiers in AI papers.
• AlphaZero Togyzkumalak - reinforcement learning agent for the traditional Central Asian game.

Publications

My research papers and publications can be found on the publications page.