I build language models and software from scratch — starting with my own language.

Self-taught builder from Bukhara, Uzbekistan. I trained a 103M-parameter Uzbek language model from an empty file, built a custom Uzbek tokenizer, and founded a learning platform that's live for 1,000+ students. I ship real things; I'm not here to write papers about them.

uzbek-gpt-103m — a 103M-parameter decoder-only transformer (RoPE · RMSNorm · SwiGLU), pre-trained from scratch on ~1.06B Uzbek tokens.

Single RTX 4090 · ~3.4 hours · ~$3.60 to train · 1.105 bits/byte — beats mGPT-1.3B + QLoRA (1.147) at ~1/13th the size.

uzbek-bpe-16k — a custom BPE tokenizer built specifically for Uzbek (handles oʻ, gʻ, and apostrophes that multilingual tokenizers fragment).

16,384 vocab · 1.839 tokens/word fertility. For a low-resource language, the tokenizer is the highest-leverage choice — that's the whole thesis.

uzbek-gpt-from-scratch — the full training code, data pipeline, and a from-scratch-vs-fine-tuning study, plus an evaluation harness for scoring Uzbek LLMs fairly.

EduBoost — a free full-stack learning platform (Next.js · TypeScript · PostgreSQL) where students teach students. Built solo, running in production for 1,000+ learners in underserved regions of Uzbekistan.

Python · PyTorch · Transformers · TypeScript · Next.js · PostgreSQL · Node.js Low-resource NLP · tokenization · from-scratch training · full-stack web · Kaggle tabular ML (XGBoost / LightGBM / CatBoost)

Bukhara → Navoiy, Uzbekistan · speaks Uzbek, Russian, Tajik, English Outside code I do traditional Uzbek beadwork (munchoq) with my family — patterns built one bead at a time on a grid. Goal: build frontier-capable language models and start my own AI company.