Eldana
Ashirova

Built a machine learning pipeline for a two-task lending challenge: predicting loan applicants’ Risk Tier using five-class classification and estimating their Interest Rate with regression. The project used 35,000 training samples with 55 original features.

Focused heavily on feature engineering, expanding the dataset from 55 to 218 features. Key features included a severity-weighted delinquency score, a synthetic FICO-style credit score, utilisation interactions, disposable income, debt-to-income ratios, and demographic percentile ranks. Out-of-fold target encoding was used to reduce leakage risk.

Trained and tuned XGBoost, LightGBM, and CatBoost models using Optuna, then combined them in a multi-level stacking ensemble. Feature engineering improved accuracy from approximately 0.53 to 0.84, and the final ensemble achieved a combined score of 0.8503, outperforming the benchmark models.

Built a machine learning pipeline to classify office buildings into five quality categories using 79 tabular features related to size, layout, amenities, zoning, construction year, and condition. The dataset included 35,000 labelled buildings and 15,000 test buildings.

Designed a preprocessing and feature engineering workflow that handled high-missingness columns, transformed year fields into age features, and created ratio-based indicators such as office area per floor, rooms per unit area, and parking efficiency. I also applied frequency encoding, out-of-fold target encoding, and log-transformations for skewed numeric features.

Trained and stacked multiple model families, including XGBoost, CatBoost, TabNet, and an MLP. The final ensemble achieved 87.37% cross-validation accuracy, significantly outperforming the 51.5% logistic regression baseline.

Starting at age fifteen, I co-authored three years of experimental research into semiconductor nanocomposite thin films, with applications in solar energy, optical sensing, and nanophotonics. The work involved designing fabrication systems, running iterative parameter tuning, and conducting quantitative characterisation using SEM, EDX, Raman spectroscopy, and optical spectroscopy.

I identified nonlinear optical behaviour and plasmon resonance effects through experimental data interpretation, and developed a predictive structural model linking material composition to optical performance. The research was recognised internationally, receiving the Best Project award at the MILSET Expo Sciences and a Gold Medal at the National Scientific Project Competition.

This experience gave me a foundation that formal coursework rarely provides: a feel for how real data behaves under experimental uncertainty, and the discipline to extract meaningful conclusions from noisy, high-dimensional observations.