Working on LLM fine-tuning and building pipelines leveraging LLMs for various use cases.

Worked on LLM finetuning comparing machine-generated datasets and human-generated datasets. A detailed report, code, and video is presented here: https://github.com/YashSharma/LLMFineTune

Worked on Computer Vision Foundation Models. Developed evaluation suite for CVML model. Product launch: https://scale.com/blog/afm1

• Worked on medical imaging (histopathology, MRI) and genomics data under the guidance of Dr. Sana Syed, UVA School of Medicine, and Prof. Donald Brown, UVA School of Data Science. Focused on domain adaptation and explainability challenges of medical imaging. • Our past work was published in MIDL, EMBC, BHI, NeurIPS workshop, AIDP, MICCAI, and MLHC.• Mentored MSDS students for their capstone project on medical imaging. Work published in IEEE SIEDS conference. • Reviewed papers for MIDL, BHI, MDPI, and multiple health conferences.

Worked as part of feature engineering and ML modeling team for developing features and learning the temporal trend in sales journey for forecasting deal closure. Used Xgboost with dynamic sample weighting for tackling the challenge of imbalance in the data. Used Bayesian optimization framework and MLflow for tuning the hyperparameters by validating performance across multiple months, achieving precision >= 0.5 and recall >= 0.4 for all the accounts.

Projects:• Clinical Trial Design Optimization: Created data mining modules comprising of XML parser, coreNLP, MeSH ontologies, CRFs, etc. for entity extraction, and entity linking tasks from clinicaltrials.gov and PubMed to expedite knowledge discovery for trial researchers. Positive response from clients and multiple use cases across therapeutic areas resulted in prototype getting converted into the proprietary product offering.• Competitive Intelligence: Predicted launch date of competitor drugs using GLM trained on event impact features along with trial features to aid the client in strategic decision making. Implemented semi-state markov model to learn trial and global event impacts in a drug journey. Achieved MSE of 10 months for phase 1 trial, 6 months for phase 2 trial, and 3 months for phase 3 trial drugs.• Information Extractor: Built a text pre-processing pipeline and implemented "Self-Taught CNN for Short Text Clustering" to extract and identify the topic of signals from Combination Product pertaining news articles and conference papers.• Safety Alert Tool: Used LDA and a hierarchical tf-idf clustering algorithm to analyze site safety logs. Trained XGBoost model on cluster matrix to alert workers about potential incidents at pharmaceutical sites. • Key Influencer Mapping: Developed an approach based on SNA using PageRank to identify KOL in the co-author network and a text-matching clustering approach for author disambiguation. Approach was accepted and presented at 22nd Merck Technology Symposium.• Response Analyzer: Used incrementally trained word vector on medical corpus along with WMD semantic matching technique to identify whether managers are complying to client's coaching standards for medical representative training or not.• Pharmacovigilance: Developed an application to identify tweets indicating adverse effects of the concerned drug using CNN model.

Projects:• Statistical Testing Module: Designed a hypothesis testing module to aid in feature engineering, feature transformation, and outlier detection.• Ensemble Modeling Module: Built a module to select appropriate bagging/boosting model based on performance metric and tune its hyperparameters using Hyperopt.