Research
Clinical AI at Cedars‑Sinai
Since August 2026 I’ve been a Research Assistant at Cedars‑Sinai in Los Angeles, building AI and machine‑learning systems for clinical research.
- CohortIQ: an LLM‑based clinical phenotyping system that combines an NLP rules layer and RAG retrieval under an AI‑agent orchestrator, with a Streamlit interface. Evaluated on a 6,110‑patient population, it cut token usage by 90%.
- Predicting pCR from breast MRI: an end‑to‑end deep learning pipeline that predicts pathologic complete response to neoadjuvant therapy, using multimodal AI and CT‑pretrained foundation models on multi‑sequence MRI and clinical notes from 1,451 pCR‑labeled patients.
Causal Mediation Analysis with Longitudinal Mediators & Competing Risks
My work in the Valeri Lab (2025–2026) focused on extending the CMAverse
R package. I have developed and implemented a user‑friendly command, cmest_pathcomprisk, which automates causal
mediation analysis when (i) mediators are measured repeatedly over time and (ii)
both mediators and outcomes are subject to competing events such as death.
Diseases of aging—e.g. dementia—often face the competing risk of mortality. Ignoring this alters estimates of indirect effects. Our command provides path‑specific effect estimates that remain causally interpretable under standard identifiability assumptions. I validated this function through extensive simulations, shipped a full testing suite and implementation guide, and applied it to 5,000+ longitudinal patient observations to study aging‑related health outcomes.
Visit CMAverse Project Website →
cmest_pathcomprisk on GitHub →
Functional Data Analysis of Mitochondrial Morphology
Under Prof. Todd Ogden, I built an automated image‑processing pipeline for 650+
biological objects (mitochondrial and egg morphology) and applied functional data analysis tools (MHD, fdasrvf)
to 3‑D mitochondrial shape trajectories in running vs. sedentary mice. We quantify shape
variability along geodesic paths, using distance matrices to relate morphological features
to metabolic phenotypes.
The figure illustrates geodesic deformation between hypothetical shapes (adapted from
Zhang et al., "Nonparametric k‑sample test on shape spaces with applications to mitochondrial shape analysis", 2022).
Avian Cranial Evolution
I worked with Rossy Natale, Prof. Graham Slater, and Prof. John Bates to examine the semi‑independent evolution of bird beaks and braincases using high-density 3-D landmarks from shorebirds (order Charadriiformes). We compared 11 alternative “modularity” models. We found that the beak has followed its own semi-independent evolutionary path while remaining moderately correlated with the rest of the skull. This research project has been presented at the 2022 American Ornithological Society Annual Conference in Puerto Rico.
Data Science Projects
30‑Day Hospital Readmission Prediction
An end‑to‑end machine learning pipeline on a 50,000‑patient synthetic dataset, with clinically grounded features (LACE score, Charlson Comorbidity Index, high‑utilizer and frailty proxies). A leakage‑safe scikit‑learn pipeline with 5‑fold cross‑validation compares logistic regression and random forest models using ROC‑AUC and precision‑recall on imbalanced data. Flagging the top‑risk discharge cohort translates into an estimated ~$620K net annual savings (62% ROI). Python, scikit‑learn, pandas.
Navi Analytics Agent
A reproducible ETL pipeline that turns raw mobile‑app, agent, and subscription logs into a curated SQLite warehouse (dimension, fact, and mart layers) with documented data‑quality checks. On top of it sit a Streamlit dashboard for DAU, retention, agent‑quality, and subscription‑conversion KPIs, and a grounded LLM analytics agent that answers natural‑language metric questions through read‑only tool calls. Python, SQL, SQLite, Streamlit, LLM tool‑calling.