Computation Student Intern, Lawrence Livermore National Laboratory (Jun - Sep 2017):

• Influence Analysis for Performance Data: Performed a novel, model-agnostic influence analysis using Graph Signal Processing to quantify parameter and data sample influence on the performance of a High Performance Computing application (Python)
• Non-black-box method whose results aligned with the intuition of HPC experts (and to be tested on more apps)
• Validated results with Neural Networks, Gradient Boosting, Random Forests (Keras, scikit-learn, TensorFlow)
• Selected to present a poster summarizing our findings and methodology at the premier Supercomputing conference SC17
• A Graph Signal Processing Approach for Sample Influence Analysis: Identified influential samples and computed image saliency maps in Kaggle and ImageNet datasets using AlexNet’s latent space (with/without fine-tuning) using Graph Signal Processing (submitted to ICML 2018 - arXiv link here)
• Presented a poster summarizing our methodology and findings at the SoCal Machine Learning Symposium 2017

Other Projects

• Facial Expression Recognition: Created a gradient boosted ensemble of Convolutional Neural Networks and a K-nearest neighbors model after reducing dimensionality using Principal Components Analysis to predict facial expressions (49% accuracy on 7-class data) (TensorFlow, scikit-learn, GraphLab)
  
• Search Engine: Built a search engine with a web interface containing modules for tokenizing, indexing, and ranking webpages in UC Irvine's website (using PageRank, tf-idf, bigrams and tag-based attributes) (Python)
  
• Compiler Vectorization Prediction: Improved the prediction accuracy of compiler vectorization by 6% using Random Forests and K-nearest neighbors after performing data augmentation with SMOTE and synthpop
  
• Rainfall Prediction (Class Kaggle project): Predicted the probability of rainfall in a location using Random Forests, Logistic Regression, Boosted Decision Trees and Neural Networks in Python (model performance in top ~20% of class)
  
• Sokoban Solver: Implemented a system for solving the Sokoban puzzle using various search strategies (A*, Iterative Deepening A*, Breadth First Search) and heuristics (such as the Manhattan distance) in C++
  
• Emotion Identification of Songs: Predicted the emotions of songs using their audio and lyrical content with Support Vector Machines, Naive Bayes, Random Forests, and word lists in Python and Weka
  

Other Work Experience:

Decision Scientist, Mu Sigma (Aug 2014 - Apr 2016):

• Devised and implemented a solution to recommend the optimal marketing expenditure allocation for an Australian insurer using multiplicative models, regression and nonlinear optimization on time series data (in R)
• Projected savings of ~7% ($2M) of annual marketing budget (recommendations were tested from Jul ‘16)
• Won the “D-WOW” award, a highly selective recognition awarded by the CEO to outstanding projects
• Awarded the “Impact Award” for showing utmost ownership, dependability and grit in all endeavors
• Developed a solution to predict fraudulent insurance claims and increase the efficiency of insurance claims investigators using Exploratory Data Analysis, Logistic Regression and clustering techniques (in R)
• Served as Teaching Assistant for our hands-on introduction to data science program (for new employees)