● Spearheaded a RAG project utilizing Chain of thought prompts and Mistral LLM to develop a robust knowledge base for client’s unstructured documents achieving 86.7% BERT score● Fine tuned a multimodal model using Llama and Vision transformers as backbone to extract fields from documents, achieving 98.5% F1 score and reducing manual data entry time by 80% as part of an Intelligent Document Processing● Built end-to-end pipelines for data loading, training, and deployment of models using Docker and Kubernetes● Designed and implemented machine learning solutions, increasing straight-through processing (STP) rates by 40% and reducing error rates by 25%.

● Utilized Apache Spark and Hadoop to gather historical data on home equity investments and trained a predictive model using XGBoost to increase the accuracy of risk assessment to 93%● Gathered user interaction data using Pandas and Numpy library from the fintech platform including user profiles, investment transactions, and engagement metrics● Worked with the frontend team to enhance user experience and increased session time by 10% using A/B testing

● Led the development of an NLP model using PyTorch to compare similarity between 75k documents, reduced the inference latency to 3.5 seconds● Designed and implemented a model monitoring and logging pipeline, tracking key performance metrics to proactively identify drift which led to an improvement in model accuracy and stability by 25%● Containerized the app using Docker, and deployed the application as a scalable Django API using AWS ECS

● Designed a computer vision based virtual online exam proctoring model by capturing facial recognition, head pose and eye gaze through webcam using cascaded YOLO models and OpenCV● Implemented real-time monitoring of over 500 users concurrently, achieving an average accuracy rate of 88% in detecting suspicious behavior and ensuring academic integrity during online examinations● Engineered DynamoDB tables for storing temporary data, resulting in efficient retrieval of items by 15%