Shashank Bemberkar

Working as a GenAI Research Engineer focused on evaluation, adversarial testing, and alignment of large-scale LLMs, including Gemini 2.5 and Gemini 3. I specialize in stress-testing models to identify hallucinations, reasoning failures, and edge-case behaviors, and in building structured error taxonomies that inform model refinement and RLHF workflows. My work involves designing adversarial prompt strategies, automating evaluation pipelines in Python, analyzing OOD failures, and collaborating closely with research, engineering, and product teams to translate evaluation insights into measurable model improvements. I also lead small research teams, ensuring high-quality data collection, review rigor, and timely delivery of insights to senior stakeholders.

Designed and implemented RAG-based GenAI architectures using LangChain and FAISS to generate contextual patient summaries from historical EHR data, reducing physician review time by 25% while maintaining interpretability. Built and deployed predictive ML and deep learning models (XGBoost, Random Forest, LSTMs, CNNs) to forecast ER visits, hospital admissions, and disease risk, improving early alert accuracy by 22–30%. Fine-tuned ClinicalBERT for clinical text classification and integrated NLP embeddings with structured EHR data, extending pipelines with LLM-assisted summarization to support physician decision-making. Deployed ML and GenAI solutions using AWS SageMaker, Docker, CI/CD pipelines, and ONNX, ensuring reproducibility, monitoring, and HIPAA-compliant production workflows. Developed scalable data ingestion and transformation pipelines using AWS Glue, Redshift, and PySpark, reducing manual processing by 40% and enabling near real-time analytics. Created explainable AI frameworks using SHAP and LIME and partnered with clinicians, engineers, and researchers to align AI outputs with clinical, operational, and compliance requirements.

Conducted exploratory data analysis (EDA), statistical modeling, and regression analysis on large-scale patient outcome and operational datasets using Python and SQL. Identified trends in readmission rates, ER utilization, and care outcomes, reducing analytics turnaround time by 20% and supporting data-driven decision-making. Built and maintained interactive Power BI dashboards tracking 15+ clinical and operational KPIs for clinicians and administrators. Performed data validation, feature engineering, and reporting to support downstream ML model development and operational analytics initiatives.

Engineered and fine-tuned ML classification models for sales prediction using Python and Scikit-learn, increasing forecast accuracy by 21%. Developed ETL workflows with Informatica and AWS Glue to automate ingestion of high-volume transactional data, reducing processing time by 25%. Built interactive BI dashboards (Tableau, Power BI) to deliver real-time KPIs and sales insights, improving decision-making speed at the leadership level. Designed customer segmentation models using clustering and NLP (TF-IDF + logistic regression) on feedback text, improving targeting effectiveness. Conducted exploratory analysis and feature engineering using SQL + Python, reducing analytics turnaround by 15% compared to SAS-only workflows. Integrated multimodal datasets (sales, financial, operational logs) to support AI-driven business intelligence reporting. Contributed to MLOps practices by containerizing models with Docker and enabling CI/CD deployments via Jenkins, reducing model rollout time.