Computer Science and Economics student at Webster University (B.S. and B.A., expected May 2027), studying at the Tashkent campus.
I train small language models from scratch and test whether they compute answers or only remember them. I also build software, and I like games.
Applying to research-based MSc programmes for Fall 2027.
MathGen: a decoder-only transformer trained from scratch on synthetic mathematics, on a 4 GB laptop GPU. RoPE, SwiGLU, RMSNorm and a custom BPE tokenizer, all written in PyTorch.
My first model scored about 80%. An audit showed that 48.8% of its validation samples were exact copies of training samples, so that number measured memorization. I rebuilt the evaluation with hash-based splits, held-out value blocks and longer operands.
Pilot results (about 5M parameters, 5 seeds):
- 99.7% on problems seen in training
- 92.0% on unseen problems in the training range
- 39.4% on value blocks held out from training
- 0.3% when operands have one more digit
These are pilot runs. The full study and report are in progress.
- NVIDIA/Megatron-LM #5445: fixed infinite recursion in the abstract tokenizer's special-token-ID property aliases.
- NVIDIA/SkillSpector #166 and #167: detection of reflective
execthroughgetattr()literals, and of Unicode Tag-block "ASCII smuggling" hidden instructions in AI-agent skills.
- alirezarezvani/claude-skills #849: proposed an i18n and localization skill. The maintainer built it as the
i18n-localizationplugin (#1095) and released it in v2.14.7.
- Three years building production web platforms in React and TypeScript, mainly for logistics companies.
- Instructor at IT-Academy, Tashkent (2023 to 2025): taught web development and ran code reviews.
- Built a math-learning web app, whose problem generators became the engine of MathGen.
Python, PyTorch, C++, C#, TypeScript, React, Git