Currently on-going Projects

Past Projects

COMPOSER Project, UCSD-UCI-UCLA

remote | Oct 2025 – Present

∙ Conducts 10+ semi-structured interviews with clinical, compliance, and administrative staff at 3 hospitals to understand potential challenges, barriers, and considerations for implementing AI-powered sepsis prediction and alert system called COMPOSER at those sites.

Inmate Violence Reports Project, University of California Irvine

Irvine, CA | Dec 2025 – Present

∙ Develops dictionary, rule, and regular expression based NLP approaches that extract key info from Arizona and Pennsylvania prison violence reports and evaluate match rates with human coded data.

∙ Applies local LLMs including Gemma and Phi for computational efficiency and preventing data leakage for key info that are more contextual and difficult to extract using traditional NLP approaches.

∙ Identifies bias or stigma in the reports via Llama fine-tuned for detecting stigmatizing language and the human coded data by analyzing disparities in severity of disciplinary action across various filters including sex, race, custody levels.

GRACE+ Prompt Engineering Project, University of California Irvine

Irvine, CA | Nov 2025 – Present

∙ Designs a new LLM prompt engineering framework that combines G(Ground Rules), R(Roles), A(Ask), C(Chain of Thought), and E(Expectations) with other methods that were effective in prior literature.

∙ Designs a research study that compares differential diagnosis and treatment plans that were generated by GRACE+ prompt against those generated by basic prompts on the same clinical cases.

Graduate Research Assistant, University of California Irvine

Irvine, CA | Jan – March 2023

Adapting the HOPE Online Support Intervention to Increase Uptake of Medications for Opioid Use Disorder

∙ Conducted phone screening of potential patients living in rural areas of California with opioid misuse or addiction issues for a Randomized Control Trial (RCT) evaluating the effects of Peer Leader Interventions for risk reducing behavior. Used a screener script that incorporated the DSM-5 criteria.