I build language models and software from scratch — starting with my own language.
A 103M-parameter decoder-only transformer (RoPE, RMSNorm, SwiGLU), pre-trained from scratch on ~1.06 billion Uzbek tokens. Trained on a single RTX 4090 in about 3.4 hours for roughly $3.60 — and it beats mGPT-1.3B + QLoRA on bits-per-byte despite being an order of magnitude smaller.
A custom byte-pair-encoding tokenizer built specifically for Uzbek — handling the language's own characters (oʻ, gʻ) and apostrophes that off-the-shelf multilingual tokenizers fragment badly. Lower fertility means fewer tokens per word, which means cheaper training and longer effective context. This is the higher-leverage half of the whole project.
A free, full-stack learning platform where students teach students — built solo end to end and running in production for over a thousand learners in underserved regions of Uzbekistan. Next.js, TypeScript, and PostgreSQL, launched in 2025.
The full training code, data pipeline, and a from-scratch-vs-fine-tuning study behind the model — plus an evaluation harness for measuring Uzbek language models fairly. The argument: for a low-resource language, the tokenizer is the choice that matters most.
I'm a self-taught builder from Bukhara. I don't learn things by reading about them — I learn by building the real thing and shipping it. That's how I taught myself full-stack development, and it's how I trained a language model for Uzbek starting from an empty file: tokenizer, data pipeline, architecture, training run, evaluation.
Alongside the software, I do traditional Uzbek beadwork with my family — patterns built one bead at a time on a grid, which is closer to how I think about systems than it might sound. My goal is to build frontier-capable language models and start my own AI company.