Abror Shopulatov

I'm an AI Engineering student at MBZUAI, working on AI systems in Prof. Abdulrahman Mahmoud's Computer Architecture Lab. Currently, I am doing PortForge-TT, an agentic system that ports pretrained models onto Tenstorrent accelerators with numerical guarantees, and previously worked on LLM-based assembly transpilation across x86, ARM and RISC-V.

I also work as a Machine Learning Engineer at tilmoch.ai, building production NLP for Uzbek and other Turkic languages: the largest open text corpora (UzBooks2, UzCrawl), SOTA Uzbek encoders (tahrirchi-bert), and the stack behind the tilmoch.ai Translator, now serving 16 languages.

Email  /  CV  /  Scholar  /  Github  /  Huggingface

Research

I'm interested in AI systems — LLMs for low-level code, accelerators, and agent infrastructure — with roots in NLP for low-resource Turkic languages.

Coding agents port pretrained models onto Tenstorrent accelerators, verified by an independent reference evaluator. All four campaigns (NLLB-200, Chronos-2, ESM-2, Parakeet TDT) pass FP32 gates; reduced-precision ports (BF16, BFP8_b) are admitted through NRMSE and task-metric gates, and profile-guided optimization delivers up to 18x speedups. Now being extended into a full paper.

Extended LLM-based CISC-to-RISC assembly transpilation to O2 optimization across x86-64, AArch64 and RISC-V, with a GCP serving and evaluation pipeline (vLLM, cross-compilation, QEMU). Led the pivot to agentic transpilation: a Harbor-based repair harness where frontier coding agents fix mistranslated assembly at near-perfect rates.

With modern techniques in Machine Translation research, we have collected dataset and trained MT model for Southern Uzbek, a language with around 5 million speakers.

I created UzLiB, a comprehensive linguistic benchmark to evaluate LLMs on Uzbek. The results revealed that even top models fail to surpass 70% accuracy, highlighting a key gap in their capabilities.

We translated FLORES+ devset to Karakalpak language. Then, we collected 100,000 pairs involving Karakalpak to fine-tune machine translation model to improve upon existing baselines.

Experience

• Built Uzbekistan's largest open text corpora — UzBooks and UzCrawl — and drove their open-source release.
• Adapted GECToR for Uzbek grammar correction, pretraining tahrirchi-bert encoders from scratch for the production Tahrirchi editor.
• Built the production machine-translation system now serving 16 languages, with LLM-based synthetic data pipelines and the Sayqalchi refinement layer.
• Co-authored WMT 2024/2025 papers on low-resource MT for Turkic languages.

Miscellanea

Timeline

I believe in the idea of "CV of failures" to better represent what I have and have not accomplished. This is my attempt to do that

• Missed the top-10 at the CHIA hackathon.
• Agentic assembly transpilation paper rejected from ASPLOS.
• Failed several hackathons in Tashkent and HackEurope 2026.
• 1st place (AI track) at the first Presidential Tech Award, with the tilmoch.ai team.
• Won regional mathematics olympiads several times, but failed IMO team selection.

Feel free to steal this website's source code.