May 2022 – Aug 2022 · 3 mos · Santa Clara County, California, United States

• Built a code synthesis model that generates code for dialog systems. Code generation approaches tend to focus on solving competitive coding based problems but real-world solutions include having to string together different APIs and function calls in a logical manner.
• Created a dataset by logging system and API calls obtained on tracing bots and their respective docker environments.
• Fine-tuned and prompt-tuned Nvidia's Megatron language model on created dataset to get a significant improvement in BART Score and human evaluation metrics.
• Created synthetic libraries to evaluate the model's ability to work with arguments, perform search-like operations and chain functions. Used this as an evaluation framework for testing and comparing with SoTA code synthesis approaches like CodeGen, CodeParrot and InCoder.
• Contributed in deploying Megatron LLM as a service internally at Nvidia for developers to test.

Mar 2020 – Aug 2021 · 1 yr 5 mos

• Worked with textual data of several forms: unstructured, noisy, multilingual, processed, scarce, etc. Derived information from this data using cutting edge NLP techniques. Worked under additional memory constraints as the solutions had to be deployed in a mobile device.
• Classification of Noisy & Multilingual SMS data: Designed a methodology to process multilingual SMS data and classify messages into a logical hierarchy. Pipeline consisted of a language detection model, a truecasing model, followed by the classification network. Special points of focus include removing systemic bias from the model and detecting salient words efficiently. The inference module is entirely on-device, thus protecting sensitive user data.
• Post-process language modelling for Scene Text Recognition: Worked on language modelling post scene text recognition for Chinese and Arabic. Also designed a teacher student network to get comparable accuracy at a much lower memory usage. Special points of focus include providing support for different variations and text orientations for each language.
• Data Augmentation for Low-resource Languages: Experimented with Variational Auto-Encoders (VAEs) on German, Hindi, and Urdu. Used sequence to sequence models for paraphrasing and augmenting data. Generated data used for training several NLP modules within Samsung.

Jun 2018 – Mar 2020 · 1 yr 9 mos

• Worked on data gathered from mobile devices. This data was used to provide meaningful recommendations and suggestions to mobile phone users in a privacy-aware fashion.
• Privacy-aware Recommender System for Mobile Applications: Provided context and sequence aware recommendations for mobile applications using temporal and geo-spatial embeddings. Ensured user's sensitive data does not leave the mobile device. (Presented at the ACM Conference on Recommender Systems- RecSys 2019)
• Predicting the Correct Case of Characters in Audio and SMS Data: Designed and implemented an efficient CNN-biLSTM-CRF architecture for predicting the correct case (upper or lower) of characters. Used as a preprocessing unit for Named Entity Recognition (NER) and Named Entity Disambiguation (NED). (Presented at the IEEE International Conference on Semantic Computing- ICSC 2020)

Jul 2017 – Dec 2017 · 5 mos · Bengaluru Area, India

• Designed a novel triple extraction process for retrieving information from research papers and technical manuals using dependency parse trees and coreference resolution. (Received provisional patent under the United States Patent Law in 2019.)
• Machine Comprehension done using Bidirectional Attention Flow (BiDAF) Mechanism.
• Coupled Tensor Factorization used for Relation Schema Induction.
• Unsupervised Semantic Parsing (USP) done to enable making uncertain inference.

May 2017 – Jun 2017 · 1 mo · Bangalore

• Worked on Natural language Understanding unit of Bixby- Samsung's voice assistant. The research goal was to review testing reports, identify utterances for which the voice assistant failed and suggest changes to the model to overcome these errors. Studied the architectural units for Automatic Speech Recognition (ASR) and Natural Language Understanding (NLU) in detail and suggested significant changes.

May 2016 – Jul 2016 · 2 mos · Pune Area, India