Jisvitha Athaluri

Implemented a Retrieval-Augmented Generation (RAG) pipeline using LangChain, Pinecone, and LLaMA models. Reduced average policy query resolution time by 38% in collaboration with claims and compliance teams. Engineered LoRA-based fine-tuning workflows in PyTorch for insurance-specific QA datasets, achieving 94% domain accuracy and reducing inference cost via quantization and vLLM optimization. Deployed model serving infrastructure with FastAPI, Docker, Kubernetes, and AWS SageMaker endpoints, supporting 500+ concurrent requests with sub-2-second latency. Established LLM evaluation frameworks to measure hallucinations and factual consistency, lowering hallucinations from 12% to 3%. Optimized hybrid retrieval (dense + sparse) to improve document precision by 27% across 500K+ indexed records. Automated ML CI/CD using MLflow and Git workflows, enabling 40% faster releases. Implemented streaming document ingestion via Spark pipelines for real-time embedding refresh and retrieval consistency

Developed a semantic search platform using PyTorch embeddings and ChromaDB; indexed 50K+ papers and cut faculty literature review time by 60%. Fine-tuned BERT and RoBERTa for Named Entity Recognition achieving 93% F1. Built RESTful inferences with FastAPI and Docker for TensorFlow-based T5 summarization models, enabling production-ready NLP services across departments. Designed cross-validation and evaluation workflows to improve reproducibility; optimized Bayesian hyperparameter tuning to increase classification accuracy by 8% and stabilize performance across folds. Containerized ML workloads and conducted A/B testing to enhance retrieval relevance by 22%.

Developed a semantic search platform using PyTorch embedding models and ChromaDB, indexing 50K+ academic papers to reduce faculty literature review time by 60%. Fine-tuned BERT and RoBERTa for Named Entity Recognition, achieving 93% F1 and streamlining metadata tagging. Built RESTful inference APIs with FastAPI and Docker for TensorFlow-based T5 summarization models, enabling cross-department ML services. Created cross-validation and evaluation workflows to improve reproducibility; enhanced Bayesian optimization to boost classification accuracy by 8% and stabilize performance. Containerized workloads with Docker and standardized deployments across compute environments. Conducted prompt engineering AB tests, achieving higher zero-shot relevance. Improved embedding strategies and retrieval metrics, boosting semantic search relevance by 22%.