Nikhil Gupta

Spearheaded the development of a comprehensive AI Agent designed to answer all coverage related user queries for an insurance policy, leveraging PDF policy documents and contracts and generating responses with citations and references. Engineered an algorithm capable of parsing PDF documents into graph structures, facilitating robust context building for Retrieval Augmented Generation (RAG) models. Built a Knowledge Graph from unstructured text using LLM and utilized it to traverse and retrieve context related to user queries. Additionally, created an LLM based agent using LangChain to answer adjuster queries based on a large pool of pdf/rtf/text documents (unstructured data). Developed an LLM based Chain to chat with multi table database (structured data) to fetch required information using queries and respond using combination of text and tables wherever required. Developed ML solutions to effectively triage claims across various divisions, resulting in a 35% improvement in claims processing time while maintaining 95% accuracy.

Built and deployed machine learning models to forecast incoming claims and estimate adjuster capacity requirements, resulting in a 25% reduction in claim backlog and 20% reduction in average claims processing time through effective resource allocation. Created machine learning models that evaluated adjuster productivity based on metrics such as claims processed, processing time and revenue generated based on types and size of claims handled and individual workload. Developed a real-time claim tracking dashboard to monitor operations and provide key performance metrics. Interacted with stakeholders to gather business requirements, translated them into technical specifications, and led a team of five to develop, test, and deploy the solutions. Built a semi-autonomous AutoML platform for ML, DL, CV, and NLP tasks with on-the-fly dashboard creation. Created Text-to-SQL engine with 87% accuracy for cross-domain natural language query to SQL query conversion. Automated data and entity extraction process of scanned legal documents by finetuning OCR, ICR and NER, reducing processing time to 15 sec per 60-page document. Built NLP models to summarise cases, achieving 98% accuracy on document classification and 0.18 RMSE on outcome prediction.

Developed a deep learning-based face recognition and attendance system and deployed it in multiple educational institutions with a real-time accuracy of over 99% running on low-end edge device.

Designed and built a miniature ground surveillance robot with face and voice recognition modules to interact with users and autonomously navigate using visual SLAM on verbal commands.

President of college robotics club; Instructor in robotics workshops.