We used machine learning to improve the reaction time for our near-infrared spectroscopy (NIRS) brain computer interface. We managed to get state of the art results to help make controlling your computer with your brain a reality. Worked with Tony Wu and Ilya Sherman. Paper

We built a gesture recognition system for the Microsoft Kinect as the first version used pose tracking instead of real gesture detection and recognition in its games. By using a probabilistic graphical model (an HMM) we were able to teach it to recognize 9 gestures (including the Mortal Kombat ones: high-kick, low-kick, high-punch, low-punch, and jump) with almost perfect accuracy using very little training data. Worked with David Philipson and Francisco Cai. Paper

Graph Cut segmentation currently (as of 2010) is the state of the art algorithm for segmenting an image into its different components. (This algorithm is used in Microsoft PowerPoint 2010) We implemented this algorithm in a few days and novelly extended it by automating the image segmentation process so that it no longer requires human input and yet produces high quality image segmentations. Worked with Evan Rosen. Paper

Built a system that can precisely detect objects in an image even if it changes shape from image to image by using a Deformable Parts Model [Felzenszwalb et al '11]. We highly optimized our system to achieve good performance even while running it on a machine severely limited compute power and minimal RAM by using novel extensions. Worked with Evan Rosen. Paper

We epxlored the problem of image classification using a family of methods based on bag of words features and by using the Spatial Pyramid Matching (SPM) and Locality Contrained Linear Coding (LLC) (currently) state of the art techniques, we were able to build a system which achieved great performance on commodity machines. Engaging knowledge from data mining, we developed a novel method of using Locality sensitive hashing to find similar pictures in this corpus of images from the computer vision challenge. Worked with Evan Rosen. Paper

We used and explored different probabilistic language models for rebuilding scrambled email messages into human readable text.By using different smothing techniques (such as absolute discounting, Katz backoff, and Kenyser-Ney for unigram, bigram and trigram models) and highly optimizing our code in order to train on over 1,240,000 sentences of the data, we were able to achieve a 96.6% accuracy on the Enron email corpus. Worked with David Philipson. Paper

In this project, we build an aligment model which suggests probabilities for alignments of words between the same sentence in two different languages. In conjuntion with a language model, our model can serve to translate one language to another. Worked with David Philipson. Paper

In this project, we explore a Maximum Entropy Markov Model (MEMM) for named entity recognition (the identification of entities - people, places, and things in the sentence). We then also learned the structure of a sentence with a Probabilistic Context Free Grammar. We heavily engineered specific features to detect the required named entities (the subject and objects in the sentence).), Worked with David Philipson. Paper

We display the rule-based system we developed for the resolution task - determining the real world entity that each noun phrase refers to. We built our engine from several new state-of-the-art coreference resolution parser techniques and experimented with three different avenues for extending these techniques. Worked with David Philipson. Paper

We used network analysis and built a model of the stock market based on the residual prices over a timespan. Using various network algorithms we found that there were different clusters of companies which correlated to eachother and had a predicatable real world relevance pattern. Worked with Tony Wu and Salik Syed. Paper

Currently you have to have a deep knowledge of computer science to automate the gathering of structured data from the Internet. A few commercial solutions exist for thousands of dollars which give you some capabilities if you are willing to put in hundreds of hours to understand their complex systems. We built a dead simple firefox extension that a 10 year old can use to obtain structured data from the web in only 1-2 minutes. It is currently implemented as a Firefox extension for maximum accessibility. Worked with Amir Ghazvinian and Sean Holbert. Paper

Newtrition is an interactive visualization that allows you to compare different types and quantities of food and explore their nutritional value. We wanted to give insight into what you eat at common places (such as Burger King or McDonalds) which do not necessarily have nutrition facts available on the wrapper. The result is a simple and highly eye-opening system which enables you to make conscious decisions about your lifestyle and food choices. Worked with Matt Bush and Sean Holbert. Paper

For the algorithms contest in which we were to write a program to find the minimum spanning tree of a graph, in whatever language and using any algorithm that we deemed as best for solving the problem as quickly as possible for a range of inputs. I used C++ and explored several algorithms including Borukva's, the Karger Klein and Tarjan randomized linear time algorithm and Chazelles non-randomized comparison algorithm.

We created and engineered several different computer vision methods to write a program to play the different types of levels in the Nintendo Wii game 'Mii', in which the player has to indentify different animated characters which are performing different actions. We used several types of feature detectors and object classifiers and heavily feature engineered our algorithms for accomplishing each of the various challeneges we were presented with. Worked with Prasanna Vasudevan. Paper

We built an object detection algorithm that uses Haar Features and boosted decsion trees to detect, identify, classify, and mark office supplies in a video. Using various forms of custom feature engineering and also SIFT, we were able to quite accurately classify objects and graphically annotate their position in the video. Worked with Matt Bush and Riddhi Mittal. Paper

We built a game playing program using a combination of diverse probabilistic and logical stategies and techniques. We harnessed cutting edge methods such as a Proposition Network Solver (logical) and Markov Chain Monte Carlo depth charges (probabilistic). Worked with Joe Marrama and Salik Syed. Paper

For my freshmen introductory seminar on Robotic Motion Planning, I built a motion planner which simulates a robot moving and attempting to get around terrain to a destination. The robot is able to plan out the best path and course of action and then travels to its destination.

We develped, created, and designed hackwalkrepeat.com for Sillicon Valley engineers to discover how to incorporate a little walking into their day and live a healthy lifstyle. We focused a lot on buildng a good user interface to appeal to our target audience; play around with it (at hackwalkrepeat.com ) and let us know what you think! Worked with Maurizio Caligaris and Maria Molfina.

Implemented a thread scheduler, the ability to run programs written by users, virtual memory, and the filesystem for the pintOS operating system. This was part of CS140 Operating Systems, a difficult rite of passage in the Stanford CS systems track curriculum. Worked with David Tran.