This was all quite a while back! Judge me harshly. Especially about the incorrect tenses.
My undergraduate thesis. Western Mongolia has experienced four of the largest strike-slip earthquakes in recorded history. The Gobi-Altai fault system is a 500km long left-lateral slip zone. The Altai fault system is a 1000km long right-lateral slip zone. Combined these fault systems accomodate 10-15% of the shortening in the Indo-Eurasian collision. At the intersection, complex kinematics have created a 150 x 100km basin, called the Shargyn Basin. I propose an explanation for the kinematics of the ongoing formation of the basin. The main source of data is a 3 month field season in Mongolia.
An MIT graduate student is developing a procedure to quantify finite strain on a microscale. On a certain surface, a sample is etched (like a microprocessor) with spaced dots. Then, on a microscope photograph, the dot positions are recorded before and after deformation. However, tracking hundreds of thousands of these dots by hand is impossible. So, I am working on an algorithm to determine the deformation function between the initial and final images by tracking the dot positions. In computer vision, this is called optical flow and is often formulated as a regularized least squares problem.
Identifying the centers (red) of the etched dots.
I worked for a while as the primary software developer for TherapyCharts, a web-based therapist records system. I developed a Python and PostgreSQL based backend and a ExtJS interface. I participated in two American Psychological Association conferences performing support and sales for TherapyCharts. TherapyCharts has now partnered with Wiley Publishing and is known as “Therascribe powered by TherapyCharts”.
In high school, I was very interested in machine learning algorithms and investigated neural networks, support vector machines and unsupervised learning. I replicated various algorithms and tested their properties at different points in parameter space on the MNIST dataset which contains 60,000 handwritten numbers. Later, I used an unsupervised learning technique (Autoencoders, Hinton et. al 2006) to improve the performance of a support vector machine model in detecting Melanoma (skin cancer). My technique had a false negative rate similar to experienced doctors. I presented some of these results at ISEF 2007, 2008 and 2009.