Jun 2019 – Sep 2019 · 3 mos · San Diego County, California, United States

• Applied Semi-Unsupervised Learning (using Variational Autoencoder) on fraud detection problem.
• Augmented research paper codebase to implement on a real-world dataset - modified the loss function, fixed bugs in gradient descent.
• Built an end-to-end pipeline to apply the algorithm on a different problem and visualize the results.
• Visualized the latent space of VAE using t-SNE plots, which helped to identify patterns in a user profile and credit history indicating fraudulent behavior.
• Matched the benchmark AUC = 0.91 (ROC Curve) of the XGBoost Model using the Supervised VAE algorithm.

Apr 2019 – Jun 2019 · 2 mos

• COGS 108: Data Science in Practice

Jan 2019 – Mar 2019 · 2 mos

• COGS 108: Data Science in Practice

Aug 2017 – May 2018 · 9 mos · Greater Delhi Area

• I worked under Ranveer Chandra, Principal Researcher, Microsoft Research, Redmond US and Prof Amit Kumar, IIT Delhi. This project was an extension to FarmBeats (An IoT System for Data-Driven Agriculture). FarmBeats is one of the ten projects on the list of Satya Nadella which inspired him in 2017. Moreover, very recently Bill Gates spoke about the importance of FarmBeats Project on his YouTube Channel.
• Implementing FarmBeats in India as it is, is not feasible due to the costs involved with sensors, drones etc. So our aim was to help the farmers in India by predicting crop yield ahead of time, using publicly available datasets like weather datasets, satellite imagery datasets made available by NASA.
• Applied Machine Learning techniques to predict crop yields for farmers in India.
• Based on rainfall data obtained avg RMSE of 900 on target values ranging from 2K to 5K.
• Utilized NASA Satellite Imagery Data to decrease the avg RMSE (error) by 45% to 500.
• I am proud of this project since it aims to tackle an important problem at hand. Its predicted that by 2050, the agricultural produce of our world needs to double to match up the population rise and feed everyone. At the same time the arable land is shrinking, water tables are dwindling. So we need to find ways to improve our agricultural produce. Moreover, from this project I learnt a lot about data analytics while tackling problems at different stages of the project. Especially due to the volume of the satellite data, I had to improve the efficiency of my code by many folds, to generate the features in a feasible time frame.

May 2017 – Jul 2017 · 2 mos · Bengaluru, Karnataka, India

• Decreased the number of utterances by finding a new domain (topic) for Bixby (AI assistant in Samsung Phones) in the utterances going to web search.
• Preprocessed the data (utterances) and used NLP algorithms for topic modeling like LDA, LSI to extract new domain (topic) from the dataset of user utterances.
• Modified the stop words removal process of Python NLTK library to improve the topic modeling.