RISHI JAIN

Investigated Agentic AI paradigms, focusing on improving the robustness and adaptability of GPT-style large language models (LLMs) using In-Context Learning (ICL) and co-training methods.

Developed and implemented non-parametric enhancement strategies that boosted LLM performance without gradient-based fine-tuning.

Leveraged contextual prompts and multi-source knowledge transfer to optimize LLM adaptability and robustness.

Fine-tuned the SOTA PARSeq OCR model for CAPTCHA recognition using a custom curated dataset of over 2 million samples, boosting accuracy from 52% to 88%.

Deployed the enhanced OCR model into production using TorchServe and Docker, ensuring scalable and efficient performance.

Developed a Retrieval-Augmented Generation (RAG) system, including an effective evaluation technique to measure and optimize its performance.

Applied computer vision and Large Scale Particle Image Velocimetry (LSPIV) techniques to analyze drone footage for flood and river velocity measurement.

Developed and trained a Convolutional Neural Network (CNN) model in PyTorch for precise velocity estimation.

Utilized extracted frames and advanced Computer Vision (CV) methods to enhance the accuracy of environmental data analysis.

Designed and implemented TS-NUC, a novel model-agnostic method for counterfactual explanation in time series classification.

Achieved superior performance on UCR datasets by leveraging LSTM-Autoencoders and latent space optimization.

Enhanced transparency in ML models by generating actionable counterfactuals for both multivariate and univariate time series, bridging the interpretability gap in black-box classifiers.