Shivam Sharmaशिवम् शर्मा

I am a Data Scientist at the Pacific Northwest National Laboratory's Physical and Computational Science directorate. I work on advancing the application of Artificial Intelligence (AI) and machine learning, particularly in the context of large-scale scientific and environmental documents.

My research experience involves application of NLP techniques for information extraction from unstructured scientific and environmental data for autoregressive models. I am also actively working on the implementation of generative AI for the development of conversational AI assistants for various real-world databases such as Livewire, ARM, NEPA Documents, etc.

I am a graduate from the New Jersey Institute of Technology, with a M.Sc in Data Science with a thesis in Crisis Informatics under the guidance of Dr. Cody Buntain

Email  /  Google Scholar  /  Resume/CV  /  Master's Thesis

My research interests included, information extraction from unstructured data, and training and evaluation of Large Langauge Models on domain-specific data.

• Description: Led the development of a specialized text corpus of over 28,000 Environmental Impact Statements (EIS), enriched with structured metadata and named entities for enhanced information retrieval. This corpus supports the development of AI-driven tools designed to improve the efficiency of NEPA reviews and aid in environmental decision-making.

• Description: Developed a benchmark dataset to evaluate comprehension of environmental review documents, comparing the effectiveness of Retrieval-Augmented Generation (RAG) pipelines against long-context modeling approaches for document retrieval and semantic understanding.

• Description: Created a domain-specific dataset for the chemistry field, curating scientific publications to enable domain adaptation of GPT-style LLMs. Trained and fine-tuned models on this dataset to benchmark performance on domain-specific NLP tasks relative to general-purpose LLMs.

• Description: Designed a hybrid feature fusion model for fake-news stance detection, integrating neural embeddings, statistical NLP features, and custom feature engineering within a deep neural network layer for improved stance classification accuracy.
• Outcome: Improved state-of-the-art performances on Fake-News Challenge 1 by 1.25% on overall score.

• Contributed to domain-specific LLM applications for scientific and environmental datasets, focusing on fine-tuning and domain embedding optimization.
• Developing AI chatbots for database websites using LLMs on AWS Bedrock, enhancing natural language understanding for seamless data access.
• Integrating seismic time-series data into Multi-Modal Models for information extraction and open-ended text generation tasks, enhancing contextual insights and data utilization.

• Supported the implementation of NLP techniques for effective data curation and post-processing tasks.
• Contributed to data extraction and processing from chemistry-domain scientific PDFs to support the training and evaluation of LLMs.

• Led research on training and evaluation of BERT-style LLMs to analyze and classify social media posts during crisis situations.

• Played a key role in research and implementation of NLP techniques for detecting stances in fake news content utilizing semantic similarity and topic modeling approaches.

I borrowed this website layout from here!