Projects
Machine Learning for Game Level Optimization
USC Viterbi School of Engineering, 2019
Under the supervision of Scott Easley, Senior Lecturer of Computer Science at USC, my team and I set out to develop a machine learning algorithm to design levels of a tower defense game. We optimized our algorithm to maximize the player enjoyment, defined as a combination of strategy (how many resources does the player have to work with?) and difficulty (how close do the enemies come to reaching the goal?). Levels were optimized incrementally with design updates based on the performance of an AI player (trained with a Q-learning model). Below is a screenshot from the tower defense game and a flowchart of the entire system.
Holographic Examination for Lifelike Motility
NASA Jet Propulsion Laboratory, 2018–2019
The purpose of this project was to develop a machine learning based approach to detect and track the motion of microorganisms and ultimately determine whether or not they are alive. Specifically, we used hand-labeled video data from a digital holographic microscope. We used a genetic algorithm to select the best combination of preprocessing, detection, and tracking methods and to tune the various hyperparameters. Shown below is a video of motile E. Coli along with output from a tracking algorithm I created. In the future, the software developed in this project will make its way on-board a spacecraft and travel to ocean worlds such as Europa in search of extraterrestrial life.
Cardiac Arrhythmia Detection from ECG using Deep NNs
NASA Jet Propulsion Laboratory, 2018
Cardiovascular disease is the leading cause of death in the U.S. and costs over $300 billion annually. Atrial fibrillation (AFib) is the most common heart arrhythmia and has a significant risk of death. Because of this, there is a need for automatic diagnosis of AFib from ECG signals. In this project, I set out to construct a deep neural network to diagnose heart conditions. The four categories of ECG signals I looked at were normal, AFib, other arrhythmia, and too noisy to tell. I constructed a 16-layer 1D convolutional neural network and trained it on over 8,000 ECGs. In the end, the network achieved 82% accuracy in diagnosing AFib with a false positive rate of 0.01.
Prosthetic Spinal Cord for Tetraplegic Patients
Senior Capstone, 2018
In order to restore mobility to tetraplegic patients, my group worked on a project to read EEG signals from the brain and interpret them as muscle movements. To first understand how the brain sends signals to the muscles, we set out to create a scenario in which people must control a prosthetic hand using unintuitive muscle movements. To achieve this, we created a VR environment and motor-controlled robotic hand. We then had users control the robotic hand by playing a dexterity-based game in VR and observed how they learned to control the hand. The project is in continual development, but the intermediate goal is to allow users to control a single finger using EEG readings alone.
Visualization of the Mars 2020 Rover
NASA Jet Propulsion Laboratory, 2017
The aim of this project was to assist the robotics group with the visualization of the Mars 2020 rover by writing rendering software and creating exporter functions. I developed three different programs for rendering and exporting 3D models of the rover and also debugged existing render software. In the future, the software will be used by rover drivers to generate command sequences. Different outputs of the software are shown in the pictures below.
Virtual Reality Motion Controller
2016–2017
The motivation behind this project was to create a hands-free VR motion controller, allowing users complete control of in-game movements using only their feet. The controller needed to be lightweight and portable, while still providing a natural and intuitive feel. A Wii Balance Board was used as the controller, which was connected to Unity through Bluetooth. By shifting their weight on the balance board, users were able to navigate complex terrain and even drive vehicles.
Heal
Hacktech, 2016
Heal provides quick and easy medical advice for the 21st century. No need to speak to a doctor or look up your symptoms online. Users can just message a professional through their phone, and they will be connected to a doctor immediately. Heal does not require internet access and will even define medical terminology using Wolfram technology.
Link
Git Motivated
Hacktech, 2016
Git Motivated is a web app that tells you how many calories you would burn and how much money you would save if you walked, biked, or ran instead of driving to a particular location. Just input your starting and ending location to see the difference you can make by ditching your car and getting some exercise.
Link
Likefai
HackSC, 2015
Likefai is a web app that suggests photos for users to post on Instagram. Using image recognition software, Likefai determines what types of photos you normally post and which ones are more popular among your followers and then suggests new photos from your album to post.