Project List

A bipedal fire-shooting robot designed and built for Columbia University Robotics Studio.

This Python-based application is designed to provide comprehensive insights into the medical device industry by accessing and analyzing data from various FDA databases. The application uses powerful, automated web scraping tools like Selenium and BeautifulSoup to extract detailed information on product problems, patient issues, company landscapes, and device-specific data.

The data is processed and visualized through dynamic plots and tables within a user-friendly Tkinter interface. This tool offers healthcare professionals, researchers, and industry stakeholders a clear and detailed view of the medical device landscape, facilitating informed decision-making and broad root-cause analysis.

This project applies supervised machine learning to predict the presence of cardiovascular disease using clinical data from 918 patients. Logistic Regression, K-Nearest Neighbors, Gradient Boosted Classifier, XGBoost, and a stacked ensemble were explored, with the top models achieving up to 95% ROC AUC on the test set. The full report details each step—from preprocessing and calibration to fairness auditing and subgroup threshold tuning—used to build the optimal and ethical models.

Led a psychiatry-independent research experiment on sex-divergent and genotypic effects during neurodevelopment on Bloc-1 deficient Pallid mice. Responsible for all sides of the project from behavioral experimentation to data analysis to scientific writing.

Designed a novel Endotracheal Intubation biomedical device in Columbia Biomedical Innovation class. Demonstrated knowledge of all sides from problem definition through solution/idea generation to prototyping. Demonstrated knowledge of regulatory reimbursement and intellectual property issues that accompany the device landscape and potential business opportunities.

Figure 1: Performance of the tumor stage prediction model. It includes feature importance rankings (top left), the normalized confusion matrix (top right), precision-recall curves by tumor stage (bottom left), and evaluation metrics including accuracy, log loss, classification report, and average precision scores (bottom right).
Figure 2: Predicted 15-year survival outcomes under every treatment combination of chemotherapy, radiotherapy, hormone therapy, and surgical type. It includes overall and stage-specific treatment recommendations (top and middle panels), predictions for a simulated patient (bottom left), and evaluation metrics, disclaimers, and future directions (bottom right).

This project applies supervised learning and survival modeling to predict breast cancer tumor stage and provide treatment recommendations based on long-term survival outcomes using data from Breast Cancer Gene Expression Profiles (METABRIC). An XGBoost classifier was trained to predict tumor stage, while a Random Survival Forest estimated survival probabilities across treatment combinations. These models were then used to simulate predictions for a new patient to determine the optimal treatment path based on 15-year survival.

Figure 1: Model bounding box predictions w/ confidence (top left), ground-truth doctor-annotated image (top right), precision-recall curves by cell class (bottom left), and image cell counts with a text table of model test results (bottom right).

This automated PyTorch Ultralytics YOLOv8 model detects and classifies cell types in microscopic blood smear images for malaria diagnosis. With red blood cells comprising over 95% of the data, the project explored addressing class imbalance using PyTorch Copy-Paste Augmentation, a minority class-prioritized scoring function, and class weighting in the loss function. Model performance is evaluated by comparing class-specific detection counts and predicted bounding box images versus the doctor-annotated ground truth, showcasing the performance and front-end prediction capability.

This project predicts multi-class adverse events caused by Drug-Drug Interactions (DDIs) using a gradient boosting decision tree model. CatBoost was selected for its ability to handle categorical pharmacological data and its interpretability through feature importance analysis and SHAP values. The model, trained with stratified 5-fold cross-validation on the DS1 dataset, aligns with benchmark results from the ADEP paper. Further work has included integrating real-time pharmacological data from three APIs and five datasets for improved accuracy while leveraging SHAP values for instance-level explanations (i.e. input: list of drugs; output: ‘30% risk of liver toxicity due to shared target pathway’).

This garage-built CNC machine is based on the tutorial DIY 3D-Printed Dremel CNC , but has been expanded for larger dimensions. Additionally, the machine head was designed to be modular, allowing for both CNC milling and laser engraving. It is controlled by an Arduino Uno running GRBL 1.1 firmware, with UGS as the G-code sender for CNC operations and LightBurn for laser engraving.

Fixed and modified a Creality Ender 6 3-D Printer. Installed Biqu H2 direct extruder print head with custom 3-D printed mount and cooling system. Installed drag chain BL Touch and RasPi Camera for monitoring and recording. Reconfigured the software to run off of a Raspberry Pi using Klipper and MainSail allowing remote control and further customizability.

Utilized the Google Design Thinking framework to create a Minimal Viable Product; a full-stack API-based medication reminder texting service to be integrated into a hospital’s Electronic Health Record. Leveraged doctor/professional interview to augment interface design and business plan presented to JPL City of LA etc.

Completed projects in Digital Electronics and Logic Systems, Analog electronics OpAmps and Filters, Electromagnetic Levitation, Interfacing Microcontrollers, Programming Microcontrollers in Assembly and C, Digital Input/Output System (Solenoid) – On/Off Control Timing Control, and Stepper Motor – Open-Loop Position Computer Control State Machines.

Completed projects in Neuromotor Control, Brain Imaging, Genomics, and AI in Medicine (Auditory and Visual).

Partnered with a MBA student in the Columbia Lab-To-Market program to draft a commercialization plan for a Biomedical Engineering Technology Accelerator (BiomedX) retinal microneedle assessing its business viability and technological feasibility. Worked closely with the engineering team to identify clinical needs stakeholder requirements and business cases to de-risk the technology.

Shelves for artists: Led Zeppelin, Kanye West, The Grateful Dead, The Beatles, Kid Cudi, Jimi Hendrix, and The Kinks.

Programmed and deployed the fast-paced iPhone game ‘Noble Ninja’ on the Apple Store. It was my first introduction to programming at 15 years old.

Wrote Matlab code that commands an embroidery machine to create the photographed design inspired by my favorite show ‘Ted Lasso’. The embroidery design possesses fractal shapes, multiple threads and thread colors, parametric fractal shapes, and quality stitching patterns.

Utilized Altair Topology Optimization to decrease material/weight and create organic designs based on expected forces and underlying safety constraints.

Wrote Matlab code to generate the SVG files compatible with the laser cutting machine. Code creates a design that includes joining slots, part holders, name engraving, and fractal/image engraving.