Machine Learning Engineer II at Tripadvisor

Fine-tuned the embedding model to embed the query and listing texts, resulting in 5% improvement in NDCG. Implemented end-to-end semantic search algorithm featuring query understanding, candidate generation and ranking steps, resulting in a 7% increase in overall search precision. Developed embedding generation and ingestion Kubeflow pipelines leveraging Ray for distributed data processing to generate dense embeddings for customer reviews, efficiently back-filling hundreds of millions of reviews and processing millions daily, with embeddings stored in a vector database. Designed a semantic search evaluation framework with custom datasets and LLM as a judge to assess quality.

AI Engineer (NLP) at WCG Clinical

Developed a co-pilot to assist analysts in updating the documents to adhere to the latest guidelines, used Retrieval Augmented Generation (RAG) to retrieve relevant guidelines for each document and suggest changes. Automated the extraction of Schedule of Assessment tables from protocol documents (PDFs) by utilizing layout detection and OCR with prompt engineering (GPT-4). Saving thousands of hours of manual work for analysts. Reduced the latency for the extraction by 80% with better system design. Executed independent sub-tasks concurrently and eliminated RAG by including metadata for the document in the prompt.

Machine Learning Engineer at ezPT Technologies

Led the text generation project, fine-tuning LLMs (FlanT5 and Llama2) using Huggingface and PyTorch libraries for medical documentation generation based on patient inputs, saving valuable time for physical therapists. Enhanced model performance by generating synthetic data through few-shot learning, utilizing public case studies and OpenAI’s API for GPT-3.5 Turbo, and collaboration with domain experts.