Research

I am a Postdoctoral Fellow in the Department of Plant Pathology at Kansas State University, where I recently completed my Ph.D. in Computer Science. My current work focuses on "teaching" machines the language of life: pre-training large-scale language models on genomic sequences to build foundation models for fungal genomes, with scalable HPC pipelines for genome-scale data processing.

My Ph.D. research centered on stance detection enhanced by large language models - using LLMs and hybrid semi-supervised methods to automatically identify and classify opinions on contentious issues, from gun regulation and vaccines to financial reports. Throughout, I focus on building robust, explainable, and practical systems that combine the power of large models with real-world constraints.

Research Areas

01Large Language Models & Foundation Models for Genomics

Pre-training large-scale transformer-based language models for fungal genomes. I am currently pre-training a foundational LLM with up to 1.3B parameters on large-scale genomic sequence data, and building scalable HPC/ACCESS training pipelines for genome-scale data processing and model optimization. This work builds on FunBERT, a Transformer model I pre-trained from scratch on fungal genomic sequences.

02Stance Detection Enhanced by Large Language Models

Using LLMs to detect and classify stances across socially and economically important domains - gun control debates, vaccine discourse on social media, and financial text (SEC 10-K filings and earnings call transcripts). I combine semi-supervised learning with LLM-generated pseudo-labels, RAG, prompt engineering, and RLHF to improve label quality and reduce reliance on expensive manual annotation.

03Explainable & Applied Machine Learning

Applying scalable, interpretable machine learning to real-world problems beyond text, including vehicle crash severity prediction using ensemble and Transformer-based models with SHAP-based explainability, and time series anomaly detection for security auditing of medical emergency access (break-the-glass) systems.

Selected Projects

FunBERT: Pre-training and Fine-Tuning Transformers for Genomic Insights

Deep Learning + Bioinformatics

Developed a custom foundation model for fungal genomes based on the Transformer architecture, enabling genomic sequence analysis and classification in fungal DNA.

  • Pre-trained the model from scratch and fine-tuned it for downstream tasks while ensuring scalability for large-scale genomic data processing.
  • Achieved over 97% accuracy in DNA sequence classification, demonstrating the model’s effectiveness in genomic analysis.

Optimizing Financial Report Analysis with LLMs and Semi-supervised Learning

NLP + Large Language Models

Leveraged LLMs as data annotators using prompt engineering and a RAG system in semi-supervised learning to improve label quality and reduce reliance on manually labeled data for analyzing quarterly earnings call reports for a private company.

  • Fine-tuned BERT-based models for stance detection using LLM-generated pseudo-labels to enhance financial sentiment analysis in earnings call transcripts.
  • Achieved average model accuracy over 85%.
Code

AI-Driven Social Media Analysis for Vaccine Sentiment

Social Media + NLP

Analyzed ten years (2013-2022) of social media data using LLMs and NLP techniques, including lexicon-based analysis and topic modeling, to examine sentiment, stance, and trends related to vaccines.

  • Implemented large-scale data crawling pipelines and conducted temporal, geospatial, and statistical analysis.
  • Published in PLOS ONE; dataset and code released openly.
PaperDataset

Commercial Motor Vehicle Crash Severity Prediction

Machine Learning + Transportation Safety

Developed and optimized machine learning models (Random Forest, CatBoost, XGBoost, Gradient Boosting, and Transformer-based models) for crash injury severity classification.

  • Designed and implemented a scalable ML pipeline for data preprocessing, feature engineering, and model training on large-scale tabular datasets.
  • Enhanced model interpretability and explainability using the SHAP framework, providing insights into feature importance and decision-making.
Paper

Time Series Anomaly Detection in Medical Break-the-Glass

Machine Learning + Healthcare Security

Implemented anomaly detection models (Gaussian Mixture Model) to identify anomalous user sessions during medical emergency access (break-the-glass) events.

  • Developed a two-tiered prediction system: a real-time fast analysis engine and a post-hoc detailed statistical evaluation model.
  • Applied statistical modeling techniques to calculate likelihood scores for session anomalies, improving post-event auditing and security enforcement.
Paper

Resources

Google Scholar - all publications and citations

GitHub - code and datasets for my projects

About - full list of publications, experience, and skills