Areas:- LLM pretraining, finetuning- Synthetic data generation

Working with Dr Junjie Hu (https://junjiehu.github.io/) on Open Domain Question Answering / Retrieval augmented NLP, Multilinguality and Large language models (LLM)

Areas:Efficient Speech Models, Transliteration

Areas:- Multi-lingual NLP (especially low resource languages)- Large scale Neural Machine Translation - Multilingual Speech systems- Efficient TransformersAdvisors - Dr. Mitesh Khapra & Dr. Pratyush Kumar

Worked on:1) Conversational QnA2) Concept mapping to different knowledge bases (Eg: "Fast heart Beat" to Tachycardia in SnomedCT)

Areas:1) Development of Low resource Speech to Text system and deployment for real-time decoding with Kubernetes.2) Low resource Machine Translation3) Conversational Voice search with Name entity Recognition for rich customer experience.4) HeathcareBot: Answering user queries by constructing and querying a Healthcare Knowlege Graph5) Enterprise search: 5.1) Document retriever + Extractive question answering models(retriever-reader models for closed domain QA) 5.2) Knowledge graph centric approach (entity extraction and linking, relation extraction and linking, building schema and KG, template based question answering)

Areas:1) Text segmentation from printed/handwritten text images, natural images. 2) Optical Character Recognition (OCR) from segmented handwritten and printed text. 3) Invoices OCR and Entity (Date, Invoice number, address, etc.) Detection. 4) Low Resource Speech to text

I worked on detecting the severity of brain injuries in CT scans using Instance segmentation.

Projects:1) Neural style transfer2) TicTacToe bot trained using self play and q-learning(reinforcement learning)

1) Leveraged Deep learning algorithms (Combination of 2-D CNN and 1-D CNN) to predict whether an image will be Viral/Non-Viral based on Image Features and Caption.2) Scraped Image data(800,000 images) from six favorite image hosting subreddits on Reddit.com.3) Trained the model by feeding a pair of images (one viral and one Non–viral) and captions to predict the relatively viral image and caption.Packages/Frameworks used: Keras, PyMongo, NumPy, Urllib, BeautifulSoup, etc

Worked on the SchoolsForIndia Project websiteDesigned the Frontend and wrote Backend code for the website's landing page, user authentication, and database design in MongoDB. Frameworks/Libraries used: NodeJS with Express, PassportJS, HTML5, CSS3