Hugging Face
BlogHi, I’m Yeraly.
I am a high school AI researcher from Aktobe, Kazakhstan. I work on multilingual and low-resource NLP, reliable detection systems, and machine-learning systems designed for real-world constraints.
My recent work includes the first-place system in SemEval-2026 Task 13 Subtask C, KazBERT, and edge-first computer-vision systems.
Research trajectory
My work began with KazBERT, an independently developed Kazakh-focused language model. Releasing it openly and seeing other researchers use it showed me how work started outside a large laboratory could still become useful to a wider research community.
I later founded DSML Young to help students exchange experiments, find collaborators, and receive serious technical feedback. A collaboration formed through that community produced our SemEval-2026 system, which ranked first in Subtask C and was published in the ACL 2026 SemEval proceedings.
Selected research
All publicationsYoungDSMLKZ at SemEval-2026 Task 13
First among 32 participating teams in Subtask C with 0.7855 macro-F1. I developed the Dynamic MIL pipeline for long code files and contributed to exploratory analysis and handcrafted features.
KazBERT
An open Kazakh-focused BERT model with a custom tokenizer and continued masked-language-model training, later cited and evaluated in independent research.
Cross-Kingdom Virtual Cells
Summer research at Nazarbayev University on AI-based architectures for predictive biological modelling across animal, plant, and microbial systems.
Building and leadership
DSML Young
I founded a student AI research community connecting young researchers across Kazakhstan and Russia through technical discussions, workshops, collaboration, and mentorship. I also raised an age-access barrier with OpenReview. The platform subsequently updated its policy, allowing researchers aged 13–17 to create profiles with parental consent and participate in scholarly peer review.
Sentra
As CTO, I work on edge-first retail analytics: local video processing, computer-vision pipelines, operational signals, and systems designed for limited connectivity.
Robotics and edge AI
Building with real hardware taught me to account for constrained compute, imperfect sensors, latency, and the gap between a benchmark and a working system.