I served as an AI/Data Science intern within the clinical pharmacology department at Genentech, Inc. To enable unconfounded analysis of drug exposure and patient response relationships, I applied and repurposed novel machine and deep learning algorithms on late-stage clinical trials data. Throughout the project, I constructed custom causal inference, interpretable ML, and visualization packages with Python, R, and big compute tools. In addition, I presented novel findings to senior leadership to guide dosage selection and patient recruitment decisions in future clinical trials. Although I can not distribute any software from the project, I am happy to share my amazing Genentech experience with anyone who is interested.

The early identification of lung lesions is very important as lesions have a high likelihood of becoming cancerous. An accurate, computer-aided diagnosis model could improve workflow prioritization, support decision making in clinical settings, and ultimately help progress large-scale screening and global population health initiatives. Towards this, my peers and I developed and tuned deep learning models for automated chest radiograph interpretation of lung lesions with high performance (>0.8 AUROC). Leveraging popular Python packages and deep learning frameworks such as Scikit-Learn, Pandas, Numpy, Keras, and TensorFlow, we also explored the implications of class imbalance and dataset cleanliness in deep learning model development.

The adverse events dataset is compiled by the U.S. FDA where reports from consumers about adverse events involving in-market products (foods, dietary supplements and cosmetics) are collected. With thousands of reports dating back to 2004 recorded in tabular form, there is a need for a visualization platform to perform exploratory analyses on this large data set. Using RShiny, tidyverse and Plotly packages, my peer and I created a reactive dashboard that allows user customization o linked graphical displays of the adverse events data. The flexible application enables different subsets, facets and groups based on victim demographic and product categories to help users observe trends in the database.

t-Distrubuted Stochastic Neighbor Embedding (t-SNE) is a popular dimensionality reduction technique for visualizing large, high-dimensional datasets in low-dimensional space. Motivated by customizable parallelizability in lower level programming languages, my peers and I implemented the first t-SNE algorithm in C to our knowledge. In addition, we parallelized our software with OpenACC directives to leverage GPU computing and achieved significant speedup compared to popular Python implementations.

To support breast tumor diagnostics, I constructed a machine learning model for malignant vs benign cancer classification. I optimized a random forest model trained on digitalized images of fine needle aspirates from breast masses. Leveraging popular Python packages such as Scikit-Learn, Pandas, Numpy, and Scipy to tune model hyperparameters, I achieved high classification accuracies (>95%) for my training and validation datasets.

During the beginning of the Covid-19 outbreak in 2020, I participated in MIT's Beat the Pandemic Hackathon. My interdiciplinary team brainstormed and prototyped a mobile application to identify essential workers likely to transmit the virus, enabling the ability to triage limited testing resources. The technology computed risk scores based on self reporting and consented GPS-based user data. I was tasked to showcase the frontend interface for clinician and patient ease-of-use using Figma. The team presented the proposed solution to healthcare and technology leaders, demonstrating the potential for this triaging technology.

Building this portfolio website was my first introduction to web development with HTML and CSS. I self-learned the programming syntax, implemented best practices, and identified resources such as Github Pages, Bootstrap and Formspree to aid my development. I learned a lot in the layout iteration process and techniques that enable sites to scale across all devices. If you have suggestions on how I can improve my site, please don't hesitate to reach out below!

As the fields of biomedical informatics and computational biology continue to grow, Python is an increasingly popular programming language in academia and industry. For my undergraduate Departmental Honors Project, my team designed bioscience-related Python modules for bioengineering students at the University of Washington. Topics such as computational protein design, neural networks and genomic sequence analysis are explored. Partnering with faculty and academic advisors, the tutorials are undergoing pilot-testing, where student feedback will guide the next iterations of curriculum development.

For my undergraduate thesis at the UW, I developd logic degradable nanogels at the DeForest Research Group . The delivery of cell and drug-based chemotherapeutics to tumors have presented major challenges in effective cancer treatment. Towards this, I designed a chemical framework for creating user-programmable nanogels that undergo programmed degradation in response to multiple environmental cues following Boolean logic. User-specified combinations of environmental inputs (e.g., tumor-presented enzymes, reducing conditions) yield material breakdown, thereby triggering therapeutic release. Through this work, I have been awarded two research scholarships and presented my studies at two professional research symposiums.

Over the course of my Data Structures and Algorithms class at the UW, I built HuskyMaps, a navigation tool for University of Washington and regions surrounding it. Internally with Java, I implemented A* Search algorithms for path finding, an Autocomplete framework for search bar text suggestions, and used data structures such as k-d Trees and Deques for algorithm efficiency. Although I am not able to share my code here due to academic integrity, building this project has been an incredibly fulfilling experience and it allowed me to sharpen my data structures and algorithms knowledge.

Protein design has become more and more computational. Here, I designed a thermostable β-Glucanase enzyme for beer brewing. Due to high temperatures during brewing workflows, there is tremendous need in developing a β-Glucanase variant that can maximize enzyme efficacy at high temperatures. With Rosetta, Chimera and intuition about protein structure and functions, I successfully designed an enzyme with such properties, with the results verified by Molecular Dynamics Simulations via Bash scripting.

Serving as a researcher in the DeForest Research Group at the UW, I assisted in engineering a 3D arteriole-capillary-venule unit with unprecedented resolution to study malaria- infected red blood cells. To process large time series datasets generated by transmission electron microscopes during blood flow experiments, I developed image and statistical analysis software in Matlab and used these computational scripts to investigate cell-vessel interactions. Throughout the project, not only did I acquire invaluable expertise in tissue engineering and infectious disease, I was also exposed to biocomputational algorithms that are widely used today. Furthermore, I was fortunate to publish my work in Science Advances.

Taking on one of the Grand Challenges in Global Health issued by the Bill & Melinda Gates Foundation, I co-founded Anophelseize. Utilizing acoustic and chemo attractants, Anophelseize is an efficient, low-cost capturing device for male mosquitoes of the species Anopheles gambiae to be used for field malaria research. Prototypes were generated with AutoCAD, 3D printing, laser cutting and other fabrication techniques. The team secured $2000+ in prototype funding and exceled in business-health competitions. This unique, collaborative project has inspired me to further my career in biomedical informatics by opening my eyes to the importance of computational sciences in global health, given the complex large datasets that transcend national borders.

As hypothermia kills twice as many people as heat does in the US, there is need in developing an active, modular, portable heating blanket to be used before emergency room arrival. Towards this, I developed a PID controlled hypothermia blanket model using LabView simulations and basic circuitry along with my peers. Our system achieved a 2 degree/min temperature increase by tuning the PID parameters.

I served as a research associate intern within the materials science and advanced platform technology development team at Illumina, Inc, a world leader in sequencing. To reduce sequencing waste and decrease genomic sequencing turn-around time, I developed and automated surface-based DNA assays that enables multi-sample sequencing on a single surface. Throughout the project, I designed quality control assessments and analyzed next generation sequencing data with in-house software. In addition, I presented research results to senior scientists, managers and directors to support patent applications and product development in the company. Although I can not disclose many details of the project due to NDA commitments, I am happy to share my amazing Illumina experience with anyone who is interested.

Although rarely used in today’s world, analog computers can be extremely useful in solving mathematical operations and differential equations. Using electrical engineering principals, I designed and built an analog computer using a combination of circuits to solve an ordinary differential equation modeling glucose transport and metabolism within a cell with my peers. Matlab modeling and analytical calculus were used to verify the analog computers’ solutions obtained from oscilloscopes.

To better understand cardiovascular rhythms, I processed EKG signals end-to-end from signal acquisition to visualization along with my peers. With voltage input from 7 electrodes on a live human subject, large, noisy time series data were generated. Using Matlab, running average filters and other signal processing algorithms were applied in order to visualize EKG metrics such as heart rate and peak shapes.