NLP and applied ML, based in Uzbekistan. I build things end-to-end — data, models, and the product wrapped around them — mostly for low-resource languages and safety-sensitive domains.
Currently: hybrid neural/rule-based systems for Uzbek NLP, and clinically-oriented ML with explicit fairness and calibration checks.
Smart Road
Grades asphalt through ASTM D6433 instead of counting defects: a YOLO11 detector over eight distress classes mapped onto the standard's taxonomy (mAP50 0.657 over 113,648 boxes), inverse perspective mapping to recover square metres on the road plane, and the deduct-value mathematics that turns densities into a Pavement Condition Index. The engine reproduces the worked example printed in the standard — 48.6 against a published 49.
PyTorch · YOLO11 · Streamlit · model · dataset
UZ-ByT5 — Uzbek Morpheme Analyzer
A hybrid system that segments Uzbek words into root/derivational/lexical/syntactic morphemes, restores dictionary forms, and explains each part. Fine-tuned ByT5 with a deterministic rule-based fallback for anything the model gets wrong.
Python · Transformers · Streamlit · model · paper
TriageGeist
A 4-model emergency severity index (ESI) ensemble evaluated for clinical safety: split conformal prediction, asymmetric under-triage cost, and a fairness audit across sex, language, age, and site on 80,000 patients.
Bio_ClinicalBERT · Conformal prediction · Fairness auditing · demo
Uzbek Operator RAG
A retrieval-augmented generation pipeline for an Uzbek telecom operator chatbot: custom BPE tokenizer, a BERT-style transformer pretrained with MLM + SimCSE, and a hybrid retriever, with no LangChain or LlamaIndex in the stack.
PyTorch · Transformers · FAISS · model
Uzbek Operator NER
The information-extraction half of the same chatbot: a BiLSTM-CRF sequence tagger for English and Russian with custom domain entities.
PyTorch · BiLSTM-CRF · model
Stenoz
A two-stage coronary stenosis detector for X-ray angiography: a U-Net trained directly on lesion masks (test F1 0.645, 95.7% lesion recall) paired with an interpretable geometric diameter-profiling layer, validated against a synthetic-injection ground truth.
PyTorch · U-Net · Streamlit · model · paper