Shradha Agarwal (PHD-Physics)

Implementing AI in education. Spearheading the development of GradeFlow™, our proprietary AI model designed to analyze and grade complex STEM subjects with precision. I lead the R&D behind our handwritten recognition (OCR) and natural language processing pipelines, ensuring high-accuracy evaluation for math, physics, and chemistry coursework. Visit: www.grade-ant.com

- Developing a comprehensive AI-driven platform that combines speech-to-text conversion and NLP for video content annotation and employs specialized deep learning models for advanced microscopy image and video analysis. - Leading a deep learning project with a specialized shift-invariant Variational Auto Encoder (VAE) model for image generation, reconstruction, and optimization through parameter fine-tuning. Conducting comprehensive evaluations using the Structural Similarity Index (SSIM) to validate VAE's image quality preservation. - Mentorship of young researchers, establishment of research collaborations, and publication of papers in reputable high impact factor journals (IEEE and EAAI). - Design and instruction of a specialized course, NE-597, covering TEM and ML models for radiation defect characterization in microscopy images. Check out some of the lectures here:https://www.youtube.com/watch?v=laAl8r-eCsQ&t=2319s. - Received UTK Start funding as Principal Investigator for research on "In-situ Characterization of Radiation Effects on Materials through Live Video Analysis with Machine Learning." - Organized the UTK-ORNL workshop on 'Deep Learning for Microscopy Image Analysis in Materials Science,' advancing research and education in the field. https://microscopyai.utk.edu/

Head of AI and ML (AI-Research Lead) at Stealth Startup (which secured funding of $9 million USD), where I lead several advanced projects specifically tailored for Large Language Model (LLM) applications. These projects include prompt engineering for optimizing LLM responses, RAG (Retrieval-Augmented Generation) implementations using data extracted from textbooks with cosine similarity, FAISS, Streamlit, and LangChain, and robust prompt injection detection mechanisms. Additionally, I am focused on developing multiclass subject multiclass subject classification system using both synthetic data and data extracted from textbooks. For data labeling, we utilized advanced LLMs such as Zephyr, Mistral, and LLaMA, comparing their annotation performance to ensure high-quality labeled datasets. Additionally, I am also leading projects on image classification systems, text extraction from images using OCR, and performance metrics for handwriting extraction from images, all designed to enhance the capabilities and accuracy of LLMs.

- Performed R&D in computer vision and machine learning. - Resolved Occlusion Challenge in Early YOLO Versions Object Detection Issues: Our custom code tackled occlusion challenges in early YOLO versions, refining bounding box measurements to boost object detection precision. This innovation enhances real-world object recognition, contributing to more accurate applications. - Implemented Mask R-CNN and Py-Tracker to track helium void dynamics in videos from in-situ TEM studies, overcoming challenges like shifting boundaries and overlapping cavities. Mask R-CNN identified and segmented voids in each frame, enabling precise tracking, while Py-Tracker followed their movement across frames.

- Physics-based ML researcher. - Integrated ML with microscopy experiments. - Performed data pre-processing, including the use of erosion and dilation techniques, and handling imbalanced datasets. - Understandood loss functions and predicting which one is best for improved boundary prediction, challenging conventional WBCE loss functions, especially for microscopy images. - Created a Region Proposal Network and applying Non-Maximum Suppression.

- Integrated experimental and computational nuclear material scientist - Conducted experimental studies on ion/neutron irradiation and gas diffusion effects in metals and ceramics, evaluating their microstructure and properties for fusion and fission reactor materials. - Streamlined data analysis in experimental studies, transitioning from manual data extraction and basic image processing to a more advanced AI-integrated approach.

- Integrated experimental and computational material scientist. - Performed physics based materials science and electron microscopy experiments. - Head of the SEM Facility.

Full-Time Researcher at Joint Accelerators for Nanosciences and Nuclear Simulation (JANNus), Service de Recherches de Métallurgie Physique (SRMP) at Commissariat à l'Énergie Atomique et aux Énergies Alternatives (CEA) Saclay, France, with a Degree from University of Paris-Sud

Admission through IIT (Indian Institute of Technology) post-graduate rank