Dheeraj Mishra

Led implementation and co-authored research addressing catastrophic forgetting in molecular property prediction, applying continual learning and refresh-learning strategies to enhance model robustness.

Designed and deployed the MTL-PORL (Multi-Task Learner - Pareto Optimized Refresh Learning) framework using ChemBERTa, integrating refresh learning with Pareto optimization to balance stability and plasticity.

Engineered multi-task/hierarchical gradient aggregation and hyper-gradient based unlearning pipelines, significantly improving knowledge retention across sequential learning episodes.

Conducted extensive experiments on BBBP, bitter, and sweet molecular datasets, achieving Anytime Avg. Accuracies up to 94.89% and Test Accuracies up to 96.86%, while reducing forgetting measures to as low as -0.0063.

Designed and implemented an efficient OCR pipeline utilizing PyMuPDF and Tesseract to accurately extract text from over 1,000 Hindi legal documents.

Developed LLM-based summarization systems specifically tailored for low-resource Indian languages, effectively processing complex legal Hindi texts.

Expanded and preprocessed a dataset of over 2,000 Hindi legal documents, significantly improving model training efficacy and diversity.

Collaborated on advanced NLP strategies, integrating open-source tools like Open Hathi, to enhance summarization accuracy in the legal domain.

Developed advanced time-series anomaly detection models leveraging PySpark and SQL, successfully reducing false positives by 20%.

Engineered robust data preprocessing and ETL pipelines, optimizing the cleaning, imputation, normalization, and structuring of high-frequency sensor data for ML models.

Deployed production-ready Apache Airflow DAGs, automating data validation, model scheduling, and monitoring, which significantly reduced manual intervention and enhanced system reliability.

Optimized feature engineering and training workflows, leading to improved model throughput and enhanced stability within resource-constrained environments.