Selected projects ordered by technical complexity, spanning image analysis, machine learning, database systems, impact analysis, forecasting, and statistical inference.
Swedish Leaf Image Analysis — Feature Extraction & Classification
Built an image-processing and machine-learning pipeline to distinguish five Swedish leaf species across 375 samples. Converted raw leaf images to grayscale, segmented leaves with thresholding, removed noise using connected-component analysis, filled binary masks, and extracted outline and color features including area, perimeter, circularity, compactness, solidity, asymmetry, RGB, and HSV statistics. Used EDA, correlation analysis, ECDFs, hierarchical clustering, contingency tables, and entropy simulations to identify the most discriminative features, then trained Random Forest and XGBoost models that achieved 98.67% and 99.20% mean cross-validation accuracy, with 100% accuracy on the held-out test set.
Python
Image Processing
PIL
SciPy
Scikit-learn
XGBoost
Random Forest
Feature Engineering
Entropy Analysis
Species Classification — Multinomial & Clustering
Used ECDFs and hierarchical clustering to identify natural bins for continuous variables, then built species-by-bin contingency tables and converted them into multinomial probability vectors. Compared null models (uniform distribution) against alternative models to test for species-specific patterns. Simulated entropy distributions and evaluated how each variable carried species information — achieving 98% prediction accuracy.
Python
Hierarchical Clustering
Multinomial Models
ECDF
Entropy
98% Accuracy
Predicting U.S. Wildfire Incidence
End-to-end ML pipeline predicting notable wildfire occurrence (≥300 acres) across U.S. counties per month using 24 years of fire records (1992–2015). Phase 1 (EDA): built a county-month panel with FIPS codes, mapped spatial hotspots, and identified seasonal patterns through visualizations. Phase 2 (Modeling): engineered lagged features to capture fire history without data leakage, trained logistic regression and XGBoost classifiers, and evaluated with TimeSeriesSplit cross-validation, AUC-ROC, precision, recall, and F1. Demonstrates the full DS workflow from raw data to production-ready evaluation.
Featured
Python
XGBoost
Scikit-learn
Pandas
TimeSeriesSplit
AUC-ROC
Binary Classification
EDA
LStore Database Engine
Built a columnar relational database engine from scratch in Python as part of a team (Data Dragons). Implemented the full LStore architecture including base and tail pages, page ranges, bufferpool management, and a record update/merge strategy. Supports core SQL-like operations — insert, select, update, delete, and aggregate — through a custom query interface. Designed for write-optimized workloads using a log-structured approach that separates stable base data from recent tail updates, with a background merge process to maintain read performance.
Python
Database Systems
Columnar Storage
Bufferpool
LStore
Page-Based Architecture
Team Project
Project Green Challenge 2025 — Impact Report
Co-authored the official data analysis report for Turning Green's 30-day sustainability challenge, reaching 1,000+ student participants across 20+ countries. Cleaned and normalized raw survey data, developed a thematic codebook for qualitative responses, and applied paired t-tests and Wilcoxon signed-rank tests to measure pre/post behavioral and attitudinal shifts. Key finding: participants who felt personally affected by climate change were 31 percentage points more likely to plan climate-positive action. Built frequency charts, butterfly plots, stacked bar charts, and choropleth maps to communicate results to organizational stakeholders and funders.
Real-World Impact
Python
Paired t-test
Wilcoxon Signed-Rank
Mann-Whitney U
Survey Analysis
Data Visualization
Thematic Coding
Macroeconomic Time Series Forecasting
Built a full forecasting pipeline in R for GDP, exports, and population time series. Applied log transformations and differencing to achieve stationarity, used ACF/PACF plots to identify ARIMA lag structures, and selected final models via AIC/BIC. Residual diagnostics confirmed model validity, and the pipeline produced 5-year forward forecasts with confidence intervals.
R
ARIMA
Stationarity Tests
ACF/PACF
AIC/BIC
Forecasting
Statistical Inference & Distribution Analysis
Five-week progressive statistical analysis comparing distributions A, B, C, and D using histograms, ROC curves, two-sample t-tests, one-way ANOVA, and assumption diagnostics. Showed that visual EDA (histogram shape, spread, center) reliably predicts formal test outcomes — A vs C achieved AUC 0.757 and p < 2.2e-16, while D's extreme variance (σ² ≈ 1,253) violated normality and homoscedasticity assumptions. Validated conclusions using Shapiro-Wilk, Levene's test, Q-Q plots, and Tukey HSD post-hoc comparisons.
R
ANOVA
ROC / AUC
Welch t-test
Shapiro-Wilk
Tukey HSD
Statistical Inference
