I am responsible for designing, developing, and deploying an end-to-end grounded response chatbot system for SoftInWay's wiki. As a result, the efforts expended by the technical support team in resolving client inquiries were diminished by over 70%. I developed this system from its foundation, which provided me with invaluable insights and a high level of proficiency in overcoming various challenges inherent in the Machine Learning/Natural Language Processing lifecycle. These encompassed the establishment of streamlined data pipelines, the orchestration of model training and evaluation workflows, the mitigation of Large Language Model hallucinations through retrieval augmentation and semantic similarity, the establishment of a scalable and distributed deployment architecture, and the seamless integration of new data to update the system, among others.
Worked on projects in Machine Learning and Natural Language Processing. Improved the F1 score of 'Biomedical Relation Extraction System' by 15% through the integration of transformer-based models, while refactoring the pipeline with the multiprocessing module to reduce the runtime by 25%. Created a real time ticket allocation application based on clustering techniques. I also contributed to improving the accuracy of an alert system by 20% through the application of various time-series techniques. By adopting a Data-Centric approach and leveraging Stanford Lab's Snorkel AI slicing capabilities, I played a pivotal role in driving Infosys to an impressive #6 ranking in the SuperGLUE benchmark, known for its rigorous evaluation of NLU capabilities. In addition, I developed autoML functionality, customized training and deployment modules, and APIs for Infosys' low-code AI platform.