YERALY GAINULLA

High School Student and AI Researcher Aktobe, Kazakhstan

Website · GitHub · Hugging Face · LinkedIn · Media & Press · Email

SUMMARY

High school AI researcher working on multilingual and low-resource NLP, reliable machine learning, and applied AI systems. Co-authored the first-place system in SemEval-2026 Task 13, Subtask C, published in the ACL 2026 SemEval proceedings. Creator of KazBERT, an open Kazakh-focused BERT model later cited and evaluated in independent research, and founder of DSML Young, a student AI research community.

RESEARCH INTERESTS

Low-resource and multilingual NLP · Reliable AI systems · AI-generated content detection · Reinforcement learning and planning · Edge AI

EDUCATION

Aqbobek Lyceum — Aktobe, Kazakhstan

Full Scholarship

RESEARCH EXPERIENCE

SemEval-2026 Task 13, Subtask C — AI-Generated Code Detection

• Ranked 1st among 32 participating teams in SemEval-2026 Task 13, Subtask C, with a macro-F1 score of 0.7855, 5.2 percentage points above the runner-up.
• Designed and implemented a Dynamic Multiple Instance Learning pipeline for long source files, preserving evidence that standard transformer truncation discarded.
• Conducted exploratory analysis and developed structural and stylometric features based on comments, indentation, identifiers, line-level variation, character entropy, and abstract syntax trees.
• Contributed to the integration of Dynamic MIL, transformer meta-stacking, and stylometric classification in the final system.
• Co-authored “YoungDSMLKZ at SemEval-2026 Task 13: MIL-UniXcoder with Meta-Stacking and Handcrafted Features for AI-Generated Code Detection,” published in the Proceedings of the 20th International Workshop on Semantic Evaluation at ACL 2026.

ACL Anthology · Hugging Face Collection

KazBERT — Kazakh-Focused Language Model

• Adapted BERT for Kazakh through continued masked-language-model training and trained a tokenizer specifically for Kazakh text.
• Built the end-to-end experimental workflow using free Kaggle GPU resources, including data preparation, tokenization, training, evaluation, and release.
• Released the model weights, tokenizer, code, data, documentation, training materials, and technical report as open research resources.
• KazBERT has since been cited and evaluated in independent Kazakh NLP research, including morphology, information retrieval, sequence labelling, question answering, cyberbullying detection, plagiarism detection, and other language-understanding tasks.
• Extended the project through downstream models and evaluations for sentence similarity, named-entity recognition, and Kazakh-language benchmarks.

Model on Hugging Face

Cross-Kingdom Virtual Cells

• Reviewed AI-based virtual-cell systems and approaches to predictive biological modelling.
• Examined applications of machine learning and autonomous agents across animal, plant, and microbial biological systems.
• Contributed to a summer research project under Prof. Yingqiu Xie, Associate Professor of Biology at Nazarbayev University.

SELECTED PUBLICATIONS AND RESEARCH OUTPUTS

• YoungDSMLKZ at SemEval-2026 Task 13: MIL-UniXcoder with Meta-Stacking and Handcrafted Features for AI-Generated Code Detection. Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval), ACL 2026.

• KazBERT: A Custom BERT Model for the Kazakh Language. Technical report and open research release, 2025. Model

HONORS AND AWARDS

• 1st Place — SemEval-2026 Task 13, Subtask C; Team YoungDSMLKZ
• Silver Medal — INFOMATRIX-ASIA 2026, AI Programming 29 medal-winning projects from 230+ submissions (~top 13%).
• Research & Impact Award — FIRST LEGO League Astana Regional Stage
• Gold Medal — Regional ICT Skills Scientific Competition, Zhubanov University; RTSD
• Silver Medal — Young Scientist Competition, Auezov University and Baishev University; Skin-AI
• Silver Medal — Regional Computer Science Olympiad, Daryn
• Control Award — Bilim Shyny FIRST Off-Season 2026; FTC Team PH0TON 34111
• 1st Place and ₸1,500,000 Prize — Startup Battles; Sentra

LEADERSHIP AND COMMUNITY

DSML Young, DSML.kz

• Founded a student AI research community connecting high school students with peers, collaborators, technical feedback, mentors, and research opportunities.
• Grew the community to 28 members across Kazakhstan and Russia, with approximately 17,000 messages during its first year.
• Organize research discussions, technical workshops, collaborative projects, and participation in AI competitions.
• Formed the YoungDSMLKZ collaboration that produced the first-place SemEval-2026 Task 13, Subtask C system and ACL 2026 paper.

GDG Cloud Almaty

EXPERIENCE

Sentra — Edge-First Retail Analytics

• Lead technical development of an edge-first retail analytics platform for physical stores.
• Build local camera ingestion, computer-vision pipelines, event processing, dashboards, and operational risk signals.
• Design systems for privacy-sensitive video processing, limited connectivity, and constrained computing resources.
• Helped the project earn approximately ₸1.5 million in cumulative competition prizes, including 1st place at Startup Battles.

FIRST Tech Challenge — Team PH0TON 34111

• Develop autonomous robot routines, path planning, hardware control, and competition software.
• Integrate autonomous behaviour with motors, sensors, physical field constraints, and imperfect real-world sensing.
• Contributed to the software and control work recognized with the Control Award at Bilim Shyny FIRST Off-Season 2026.

American Corner and Makerspace Aktobe

• Presented Skin-AI through the Makerspace Residents Program, discussing its architecture, applications, safety considerations, and limitations.
• Taught introductory machine learning to students with different levels of programming and mathematical preparation, covering classical machine learning, neural networks, and transformers.

SELECTED TECHNICAL PROJECTS

AlphaZero for Togyz Kumalak

Python · PyTorch · Reinforcement Learning · Monte Carlo Tree Search

• Adapted and extended an AlphaZero implementation for Togyz Kumalak, implementing the complete game environment, including legal actions, captures, kazan scoring, tuzdyk rules, and terminal conditions.
• Trained a residual policy-value network through self-play and Monte Carlo Tree Search, with batched evaluation and mixed-precision training.
• Built an interactive interface allowing users to play against trained agents.

ChronoSort — Edge AI Archaeology Assistant

Python · PyTorch · SigLIP · Streamlit · Raspberry Pi

• Developed a computer-vision system for recognizing and sorting archaeological objects using a vision-language model for zero-shot classification.
• Optimized inference for Raspberry Pi 5 and integrated a real-time interface, voice feedback, GPIO control, and automated sorting mechanisms.
• Received the Research & Impact Award at the FIRST LEGO League Astana Regional Stage.

Skin-AI

Python · PyTorch · ConvNeXt · ONNX Runtime · Gradio

• Built an educational dermatology image-classification application covering 22 categories of skin conditions.
• Implemented ConvNeXt inference, confidence reporting, optimized ONNX inference, and an interactive Gradio deployment.
• Included explicit safety messaging clarifying that model outputs are informational and do not constitute a medical diagnosis.