Data Science & Machine Learning
I love mining data and the possibilities hidden underneath!
Major interests include computational linguistics, machine learning, deep learning, and climate science. At times I write for my personal blog as well.
Worked in a team of 5 and developed a pipeline for building knowledge graph out of unstructured text documents. Completely owned research & development of building a natural language query system for fetching information out of this knowledge graph
Explored how close or far is an earth system from real world physics specifically when it comes to predicting surface pCO2
Implemented physics-guided RNN as discussed in the research work by Jia et al, 2019 wherein a physics related loss is added in addition to the usual RNN loss to ensure that the predicted outputs agree with energy conservation laws. Bayesian optimization was used for fine-tuning models
Analysed relationships between different hurricane characteristics and performed feature engineering. Ultimately built a linear regression model to predict the economic loss caused by hurricanes
Implemented complex-coefficient spectrally parameterized convolution filters as well as a mechanism to perform Fourier convolutions using Keras
Led a team of 5 through this end-to-end applied ML project that involved exploratory data analysis, data cleaning, feature engineering, model selection, and model training
Analyzed data on business applications provided by DCA to understand the distribution of various businesses in the city as well as identify any disparity in the issuance of licenses specially amongst marginalized neighborhoods
Experimented with various statistical (GMM-HMM), machine learning (logistic regression, SVM), and deep learning (DNN, CNN) models as well as different representations for audio signals while building a Speech Emotion Recognition system. Achieved best in class accuracy (~85%) with DNN architecture when applied on audio signals compressed using PCA.
Built a CNN-regression model for detecting bounding boxes around independent objects in images (accuracy ~87%). Qualified till 3rd round (pre-final) in the contest.
Built a digital geospatial crime mapper and data visualization tool using R-Shiny to provide police officials with a comprehensive overview of the crime trends prevailing in their jurisdiction in real time.
Worked on this project during our participation in the EXL EQ 2018 Contest. Were provided with the data of a shopping center based in San Francisco and were asked to draw insights from this data and recommend ways of increasing the revenue for the center.
