Note: Best viewed on Chrome, Safari or firefox

Short Bio


Ganesa Thandavam Ponnuraj is a second year Computer Science Masters student at the Stony Brook University. His interests lie in the intersection of algorithms, machine-learning and natural-language-processing.

He works with Prof. Yejin Choi for his Masters Thesis.

Also, he works as a Research Assistant with Dr. Minsu Ha and Prof Ross Nehm at the Nehm lab on the evograder project


Spring 2013

Fall 2013

Spring 2014


  • EvoGrader Poster at SABER 2014
  • Masters Thesis , successfully defended on December 8th 2014
  • Select Projects


    • mod_script_nonce on github
    • An apache web-server module that provides support for script-nonce attribute proposed in the content security policy draft 2.1
    • script-nonce are associated with javascript elements on a page
    • the same nonce is sent by the web-server to the client via a special HTTP header namely Content-Security-Policy
    • the client matches the nonce in the CSP header with those in the javascript elements on a page
    • effective way to thwart stored / reflected XSS attacks

    Suite of pacman AI agents

    • Uninformed and Informed search strategies for pac position search problem in prolog and python
    • Adversarial search: min-max search and alpha-beta pruning
    • Markov decision process and QLearning agents for the Pac-man
    • Decision Tree and Naive Bayes classifiers for Hand-written digit recognition
    • Prolog pac-man agents on bitbucket
    • We were given Berkeley's pac-man framework for our projects. As per the license agreement, solutions are not posted online for classification and python pacman agents

    ESA based reverse dictionary

    • Reverse Dictionary on github
    • Explicit semantic analysis based reverse dictionary using wikipedia dump (44GB as of April 2013) as the world knowledge
    • Uses TF-IDF, cosine similarity and centroid-classifier
    • Used python for data-processing and built a database of wikipedia concepts
    • php based simple web-app for serving relevant results based on the user input
    • Built required non-clustered indices on the data-base to optimize response times

    Shadow Boundary Detector for outdoor images

    • Shadow Boundary Detector on bitbucket
    • Used 4 sets of visual features as described in Work by Huang et al.,
    • Features are extracted in 3 different scales and so we have a 36 dimensional feature vector
    • Used MATLAB and LIBSVM to train the classifier
    • Performed grid search of the regularization parameter and gamma for the RBF Kernel of the SVM
    • Classification accuracy: 88.77% F1 Score: 89.02%. Bettered the accuracy reported in the paper which was 81.63%