I train multimodal language models that see, hear and read, and I build things people can touch: Kazakh AI, edge AI and walkable 3D worlds.
A 0.6B model on the phone writes a confidence score before it answers. When it is not sure, it hands the question to a larger model in the cloud. It learns this with a self-evolving training loop over 14 rounds in seven domains, including Kazakh culture.
Kazakhstan's first language–audio–vision model, trained on more than 3,000 hours of audio. I added audio-model support to vLLM and ms-swift.
Training and ablations for a VLM on more than 10 million images and 50 million question–answer pairs, Kazakhstan's largest AI dataset.
A compact 4.3B Kazakh–Russian–English vision–language model (InternViT-300M + Qwen3-4B). I worked on training, ablations and the four-stage adaptation pipeline.
Whisper-based models fine-tuned on a 100k-hour English–Kazakh dataset. Also TilSync, a real-time subtitles app I designed and built end to end.
Archery and blade games inside a reconstruction of our university atrium. 21 releases in 16 days, played by visitors at NU Open House 2026.
Research on world models and 3D scene synthesis, plus a 7-module interactive course on Gaussian Splatting that runs in the browser.
/magical_tablet
Learn 3DGSInteractive Gaussian Splatting course./3dgs/learn
ISSAI VRWebXR games in the NU atrium./issai-vr
fogswarmOne giant model or 500 small ones in browser tabs?/distributed_agents
GS-PlaygroundA visual explainer./gs_playground
VQ-VAE from zeroThe image tokenizer, explained./vqvae
GaussianGPT from zeroTransformers that make 3D scenes./gaussiangpt
Lead of the AI Engineering track: school students build their own projects with AI agents.
“RYŌTENKAI: One-Shot Domain Expansion for Walkable 3D Worlds”, 1st International Conference on Mathematics and AI, Nazarbayev University.
ISSAI AI Notebook and ISSAI VR games at the ISSAI booth.
Designed and taught the track on training small Kazakh/English/Russian LLMs from scratch.