• Proposed a novel problem Time-synchronous multi-device system (TSMDS) that is prevalent in the real world but hasn't been unexplored
• Introduced Group supervised learning (GSL), a novel framework solving TSMDS by extending contrastive learning to multi-device setup increasing the complexity manifold
• Surpassed the Fully-supervised and existing Self-supervised baseline by atleast 0.15 F1-score utilizing a fraction of labeled data

• Speech recognition platform for voice impairments through wafer-thin, battery-free and stretchable RFID Tattoos
• Collected own sensor data and implemented Random Forest model calibrated on the stretch of tags to achieve state of the art 86% accuracy on a vocabulary size of 100 most common English words

Automated E-commerce Question-Answering system
               Flipkart, Summer 2020          

• Generated synthetic queries from a limited set of user query to increase the dataset size by more than 30%
• Combined the BERT and GPT-2 models for developing a target product-type classification system which would then prompt the text-generation model to answer user query in natural language all in real-time
• System might be implemented to the company’s website, subject to quality check approval

Deep Neural Matching Models for Graph Retrieval from Product Graph
InfoLab, Indian Institute of Technology Bombay, Autumn 2020

• Developing a novel DNN approach for searching a subgraph from a corpus of sample graphs by scoring random walks in their product graph
• Working on decreasing the query time to sub-linear time by employing hashing of corpus graphs

Answer-type tagging through constant embeddings
InfoLab, Indian Institute of Technology Bombay, Spring 2020

• Improved the performance of an existing Answer-Type Inference (ATI) RNN based manuscript by almost 5% through conducting experiments in 10 different settings across 4 QA datasets
• Outperformed the previous RNN system by 20% across the 4 QA datasets by deploying BERT architecture in ATI task

Abstractive Text Summarisation Indian Institute of Technology Bombay, Autumn 2019

• Implemented an extractive text summarizer using the page rank algorithm, achieved a ROGUE score of f-32%. Latent Semantic analysis of the generated summary and the ground truth revealed an overlap of 79%
• Comparatively, the abstractive summarizer built using seq2seq model with attention mechanism, achieved a ROUGE scores of only f-13%

Nuclear Survey
(Paper accepted at DAE-BRNS Symposium on Nuclear Physics, 2019) Tata Institute of Fundamental Research, Summer 2018

• Created robust scripts for parsing nuclear values like energy levels, spin and half-life from the international atomic nuclear dataset that helped in standardizing the nuclear levels of a vast number of known nucleus

Secure Personal Cloud Indian Institute of Technology Bombay, Autumn 2018

• Implemented a cloud based file storing system for Linux system with block level customizable file encryption schema allowing multiple clients to decrypt their data at real time simultaneously or share the same
• Developed a bash interface along with a web client for an interactive user experience