Sklearn • PyTorch • Streamlit • Jan 2025 – May 2025 • Personal project
Built a multimodal ICU outcome predictor combining LSTM vital sign modeling with BioClinicalBERT embeddings of clinical notes.
Implemented TF-IDF feature extraction from 5,000+ clinical note features with logistic regression as a baseline and improved AUROC by 6–8% over NLP-only models.
Designed end-to-end ML workflows and deployed the model as a REST API service with CI/CD integration for reproducibility.
Developed and maintained a detailed extract capturing medication dispenses at eligible facility locations, classifying each dispense (Inpatient, Outpatient, Medicaid/WAC) to ensure 340B compliance.
Modeled 25 core dispense data points and 14 auxiliary attributes (dose, strength, volume, quantity, 340B quantity) to support Verity processing.
Delivered encrypted .txt files via MoveIt and exposed the extract through a reporting portal for ad-hoc execution, saving ~114 hours/year and driving more than $12M/year in pharmacy drug spend savings.
CLABSI Audit Report | Central Line–Associated Bloodstream Infection
Oracle Health • Jan 2025 – Mar 2025
Built a report providing comprehensive information on patients with central lines and associated continuous infusions to support CLABSI auditing and patient safety.
Captured 18 critical data points including line type, dressing activity, number of lumens, infusion status per lumen, site care, and cap change timestamps over specific clinical lookback windows.
Implemented a new query design that allocates and deallocates memory within subqueries and stores data in a 2D array for efficient PDF generation and structured output.
HIDI Extract | District of Columbia Hospital Association
Oracle Health • Sep 2024 – Dec 2024
Developed a data collection and reporting solution for UB-04 claims under DCHA requirements using structured record types (A–G) based on HIDI specifications.
Orchestrated one driver script controlling six modular scripts with a shared data structure to ensure modularity and maintainability.
Captured 70+ data points with <2% error rate across patient, charge, insurance, and clinical data and delivered standardized files quarterly to an SFTP site.