Fine-tuned LLMs (Llama 3, Mistral) using LoRA & QLoRA on custom Uzbek-language datasets, improving downstream task accuracy by 30%+. Designed end-to-end STT/TTS training pipelines on 500+ hours of collected audio data; applied bitsandbytes quantization to reduce GPU memory by ~40%.Containerized all ML services with Docker & Docker Compose; deployed scalable inference APIs using FastAPI on GCP and AWS. Implemented MLOps workflows: experiment tracking, model versioning, and automated retraining pipelines for continuous improvement.

Built ML models (XGBoost, LSTM, Random Forest) for Dust Storm Classification, Groundwater Time Series Forecasting, and Air Quality Index Prediction. Performed feature engineering on geospatial and time-series datasets; improved prediction accuracy across all three tasks significantly. Delivered reproducible model training pipelines packaged with Docker for easy deployment by the client team

Created interactive live dashboards (Plotly, Power BI) with direct database connections hosted on Cloud VPS, enabling real-time business monitoring. Automated data refresh pipelines using Python scripts, reducing manual reporting effort by 80%.

Developed a GPT-powered MCQ (Multiple Choice Question) generation system in Uzbek, automating question creation for thousands of students on educational platforms. Built data collection, cleaning, and annotation pipelines; coordinated with annotators to maintain high dataset quality standards