Publications

A novel hybrid framework that integrates reinforcement learning-trained search agents with reasoning-enhanced clinical prediction models. The system replaces static knowledge graphs with dynamic RL agents that learn optimal retrieval policies for medical RAG applications, demonstrating superior performance on mortality and readmission prediction tasks using MIMIC-III/IV datasets.

A comprehensive study evaluating the alignment between lexical models and cognitive-appraisal models in measuring happiness. Using Reddit survey data on short-term and long-term subjective well-being, this research compares self-reported happiness ratings with computer-generated ratings from large language models and lexical sentiment analysis methods, revealing weak but significant correlations and highlighting the limitations of computational approaches in predicting user-reported wellness measures.

A research study exploring the adaptation of the CacheBlend framework for efficient key-value cache management in code dependency systems. The project leverages mutual import relationships between Python packages to selectively blend caches, implementing advanced eviction strategies including LRU, LFU, Least Weighted, and Composite Cache methods to optimize dependency management for faster retrieval while maintaining optimal performance under memory constraints.

A comprehensive reproduction study of "Hurtful Words: Quantifying Biases in Clinical Contextual Word Embeddings" focusing on gender bias in clinical BERT models. This research reproduces key methodologies using the Baseline Clinical BERT model and MIMIC-III dataset, implementing bias quantification through prior-adjusted log-probability scores and downstream in-hospital mortality prediction tasks. The study reveals context-dependent gender associations and measurable fairness gaps, supporting the fundamental conclusion that clinical language models exhibit quantifiable biases with critical implications for healthcare AI deployment.

An empirical analysis examining the 85% decline in telephone operator employment from 2002-2021, applying skill-biased technological change theory to understand how automation affects specific demographic groups and creates new opportunities in complementary sectors.

An experimental study investigating the relationship between gold nanoparticle size and optical properties using UV-Vis spectroscopy and TEM imaging. Research demonstrates how varying HAuCl4 concentrations affects particle diameter and wavelength characteristics for potential biomedical applications.