• Built an open-domain, contextual, conversational QA agent using modified DrQA, BERT and PGNet architectures
• It has a document retriever, neural reader, clarifying system and an answer ranking module
• Trained an attentive stacked biLSTM model having 45% F1 accuracy on QuAC dataset. Used attentive history embeddings for context
• Improved the retriever accuracy (% match in top 5 doc) from 11% to 64% by adding previous QA pairs and topic as context in queries to retriever

• Built a framework that generates questions, using transformer based uniLM model, on which a given QA system fails.
• Trained a 2 layer neural network classifier on failed questions (using infersent embeddings) of a given QA system (BiDAF, BERT based) to classify above generated questions as unanswerable/answerable, with 65% accuracy

• Developed a backup file system, BKPFS, using a stackable file system WRAPFS
• BKPFS creates a hidden backup directory and multiple backup files per file to backup any modifications to the file. Users can use ioctls to list, view, recover or delete any of the backups.
• Wrote ~2000 lines of kernel code to implement system calls, ioctl commands, visibility and retention policies.

• Detects Password reuse.
• Detect the entering of passwords on the wrong website
• Warns users on clicking links to non-Alexa 10K websites
• Used Chrome’s runtime and webNavigation API’s. Stored SHA-512 encrypted passwords in chrome storage.

Do popular songs endure?

Popular Dataset Contributions on Kaggle

• Proposed and tested 5 hypotheses (Ex. black cars value > blue cars) using KS, Wald’s & Permutation Tests
• Estimated Craiglist traffic trough time series prediction of no. of ads posted using EWMA & Auto Regression