My Portfolio

Experience

Worked on multivariate data analysis of particle interactions from CMS RunII datasets. Developed multiple data visualization and event modeling packages (C++, python) for RunII Legacy results, in collaboration with 8 scientists.
Measured the 95% confidence interval of observing the Large Extra Dimension model, and rejected the parameter of interest Lambda by using a likelihood-ratio hypothesis testing.
Further improved the sensitivity to new phenomena by introducing new angular information, and enhanced modeling robustness by 9.1%.

I built a time-series forecasting model on a company's stock prices using a TensorFlow-based Recurrent Neural Network with LSTM or GRU cells. The update gate of GRU allows one to predict future stock prices based on past data trends. The prediction precision on future stock prices reaches 0.7.
I also improved the model by using a Deep Reinforcement Network to mimic user's buy, stay or sell actions. The Q-value of DQN network can learn the reward from virtual predict-and-trade, which enables the network to learn to gain more rewards.

In order to classify brand accounts from normal twitter user accounts, I built a Classification model on user tweet dataset. With word2vec representation of word embedding and Logistic Regression, brand accounts can be easily separated.

I built a Convolution Neural Network with Conv2D, Dropout and BatchNorm layers to detect hand-written digit patterns (corners, lines), using TensorFlow based Keras on Jupyter-notebooks. The precision of CNN on these digits can reach 99.5%.
Alternatively, I also used t-SNE technique to try clustering images (since image size is small here). It also turns out to be a good method to cluster image patterns.

This is a Computer Graphics course project, based on C++ in Visual Studio.
I've built an interactive 3D graphics system consisting of a planar pinhole camera and a view constructor. The objects (teapot) are modeled with triangle mesh that stores shared vertices and triangle connectivity data. Scenes are created by rendering objects' texture and lighting with bilinear interpolation lookup and phong shading model.
I also implemented environment mapping of distant geometry with a cube-map class. This enabled the specular reflection (reflective geometry on teapot's surface) using per-pixel reflected ray lookup.
More advanced fun techniques, such as projective shadow mapping, first-surface refraction, or GPU shaders are also available.

I led a team of 4 in developing an interactive platform for incoming students to gain more comprehensive information on cross-department researches at Duke.
I used Python Scrapy to crawl down research group information from 20+ department websites, and preprocessed data with Python before parsing them into a self-defined database.
We developed the website UI interface via Html/Css, JavaScript, which allows users to choose from any department of interest.