At CVS Health, I’ve been focused on building real-world AI systems that directly improve patient experience and clinical efficiency. I led the development of a healthcare conversational AI assistant powered by LLMs and RAG pipelines using LangChain and OpenAI APIs. This system enabled accurate, context-aware responses and improved patient query resolution while reducing manual workload for support teams.

One of my key contributions was designing a semantic search layer using vector databases and embedding models, which improved information retrieval relevance by 30% across large healthcare knowledge bases. I also architected scalable microservices using FastAPI and deployed them on AWS, ensuring high availability and real-time inference performance.

On the data side, I built PySpark-based pipelines with AWS Glue to process large volumes of structured and unstructured healthcare data, enabling robust feature engineering. I also implemented prompt engineering and fine-tuning strategies that reduced LLM hallucinations by 25%.

To ensure reliability, I introduced MLflow-based monitoring and evaluation frameworks, enabling continuous model improvement and compliance tracking.

t LexisNexis, I worked on building scalable machine learning solutions to extract insights from complex legal documents. I developed an end-to-end document intelligence platform that automated classification and information extraction, significantly reducing manual processing time for legal teams.

I built and optimized machine learning models using Scikit-learn and XGBoost, and later enhanced performance by integrating transformer-based NLP models from Hugging Face—improving classification accuracy by 18%. My work involved designing robust preprocessing and feature engineering pipelines using Pandas and NumPy to handle highly unstructured legal text.

To scale the system, I engineered distributed data pipelines using PySpark, enabling batch and near real-time processing of large legal datasets. I also deployed models as REST APIs using FastAPI and Docker, reducing response latency by 25% and improving integration across internal platforms.

Additionally, I implemented MLOps workflows using MLflow and CI/CD pipelines, which streamlined experimentation, versioning, and deployment. Continuous monitoring and retraining improved model reliability by 20% over time.