Landmines are typically buried in non-homogeneous soil and at different depths. This causes the signal measured at the sensor to be warped thus introducing great variability not related to anything fundamental about the landmine shape or energy profile were these conditions not present. These latent warps hide the true statistics of the landmine profile which is required for a statistical model to properly distinguish between targets and other clutter. My research is focused on this pre-processing step- efficiently removing warps so that the statistics can be learned from the data with high fidelity.

Finding such similar regions in images, called image segmentation, is often a first step for higher level tasks such as object recognition. In an image, often, very many pixels are devoted to describe a part of an object which in fact could be summarized by the knowledge of the boundary of that region. In my work I explored different feature spaces and similarity metrics to obtain homogeneous regions at different scales in the image. To accomplish this task I had to draw on my courses in graph theory, wavelets, digital signal processing and wavelets.

In my NLP class I wrote an algorithm that could classify news articles based on their content. This used the well-known TF-IDF feature space that assigns importance to words based on their apparent importance. I also wrote a program to identify to which author a certain text passage belongs. This was done by analyzing the synonym sets an author is wont to use and thereby learn the author's style.
In my ASP class I worked on a program that could identify the sound of different events such as a mouse-click, cough, train arriving in recordings of different settings of real life. This was done by finding salient points in the spectrogram of the audio clip and using a Hidden Markov Model to characterize each type of noise.
I also took an AI class where I wrote a checkers program which you can play here. It uses an implementation of alpha-beta pruning of the game tree along with my very clever heuristic function to quantify how good a certain move is.

An insurance company has many liabilities (the policies it issues) that it must match with assets (stocks, bonds, etc) to ensure its ability to pay off the policies. The model determines whether the given set of assets, under different interest rate scenarios over time, will balance out against the liabilities. I worked on a clustering algorithm that can summarize the assets while maintaining model accuracy. This reduced the run-time of the model to a few hours down from a few days.

Identifying idiosyncrasies of a particular object is done most easily when it is represented simply. Applying what I previously learned in my real analysis classes at Cooper Union, we studied Ricci flow and conformal mappings to transform an arbitrary 3 dimensional object into a smooth sphere, without losing the important features of the original object.

One day I decided that I wanted to know whether someone read my email. I could not find such a service, so I made one.
Out of frustration at having to answer people's posts on Facebook, I made an automatic wall post generator that uses your previous posts to the given person to generate a personalized message.
There is simply too much news to read, so I made a program that identifies index words for news articles such that I can browse through those and then choose the articles I would like to read.
I wanted to learn how to program in Android, so I worked on an application that can find a suitable time for two people to meet without them having to share their schedules with each other.
...