A selection of machine learning, AI, and data science projects spanning agentic systems,
production ML, and applied econometrics.
Analyzing Demographic Biases in LLM Essay Grading
2026 · CSEN 346, Santa Clara University
2nd Place — Best Educational Impact Project
Investigated demographic bias in transformer language models (XLNet, RoBERTa, Longformer) used for automated essay scoring, measuring score disparities across gender, race, English-language-learner status, socioeconomic status, and disability.
Benchmarked across two large-scale datasets — PERSUADE 2.0 (~26K essays) and ASAP 2.0 (~25K essays) — quantifying bias with weighted standardized regression z-scores.
Implemented two mitigation strategies — adversarial debiasing via a Gradient Reversal Layer and orthogonal projection of demographic directions — finding that neither reduced bias without degrading scoring quality (QWK).
Built an autonomous AI agent leveraging OpenAI LLMs, the Gmail API, and the Google Sheets API to track, analyze, and update job-application data from emails in real time — capable of handling 1,000+ emails daily.
Engineered a semantic extraction pipeline achieving 100% accuracy in identifying company, role, recruiter, and application-status details across diverse email formats.
Reduced manual email-tracking time by 95% through end-to-end CRM synchronization, enabling a 24/7, zero-maintenance system.
December 2024 · Sacramento Municipal Utility District (SMUD)
Optimized model precision by 10% through hyperparameter tuning with Bayesian Optimization and HyperOpt.
Achieved a 20% accuracy gain by developing a K-means clustering pipeline for feature engineering, processing over 7,000 geospatial records.
Improved data quality by 100% using the ArcGIS Geocoding API to recover missing geospatial data, ensuring robust analysis.
Raised the model's decision rate by 40% by resolving training-set imbalance with SMOTE oversampling.
PythonRXGBoostK-meansArcGISDockerAirflow
Impact of Education & Government Spending on Unemployment
November 2024 · Econometrics Research
Conducted a panel-data analysis using government expenditure and unemployment data from World Bank Open Data.
Built lagged econometric models on 2,599 data points in Stata to investigate the causal relationship between education spending and unemployment, selecting optimal lags via AIC/BIC.
Found no statistically significant causal relationship, highlighting the challenges of omitted-variable bias and endogeneity.