– Built and deployed a sophisticated QnA Chatbot using LLMs and RAG, achieving a 93% user approval rate. Integrated real-time fund data and streaming, boosting user engagement by 65%.– Optimized data ingestion latency for PDFs/URLs by 4x via asynchronous processing, accelerating data indexing and vectorization, and substantially reducing ingestion time.– Implemented hybrid/semantic search, increasing context recall score by 9%, and developed a ReAct prompting-based agent for structured… Show more – Built and deployed a sophisticated QnA Chatbot using LLMs and RAG, achieving a 93% user approval rate. Integrated real-time fund data and streaming, boosting user engagement by 65%.– Optimized data ingestion latency for PDFs/URLs by 4x via asynchronous processing, accelerating data indexing and vectorization, and substantially reducing ingestion time.– Implemented hybrid/semantic search, increasing context recall score by 9%, and developed a ReAct prompting-based agent for structured data querying, significantly diversifying the bot capabilities.– Improved ticket resolution F1-Score by 7% by implementing a dynamic few-shot learning mechanism, leading to more accurate and efficient issue resolution. Show less

– Implemented a unique pretraining strategy for Extractive Question Answering using BERT, leveraging spans over start and end tokens, which led to an 8.5% enhancement in the IoU Score.– Streamlined Tabular QA engine latency by incorporating the TAPAS model, thereby eliminating the requirement for intermediate query generation and improving overall efficiency.– Developed a Post-Call KPI bot to track various key performance indicators, enhancing the post-call analysis and monitoring by… Show more – Implemented a unique pretraining strategy for Extractive Question Answering using BERT, leveraging spans over start and end tokens, which led to an 8.5% enhancement in the IoU Score.– Streamlined Tabular QA engine latency by incorporating the TAPAS model, thereby eliminating the requirement for intermediate query generation and improving overall efficiency.– Developed a Post-Call KPI bot to track various key performance indicators, enhancing the post-call analysis and monitoring by utilizing seq2seq models. Show less

– Created an Automated Invoice Detection Application to detect, process, and store the standalone invoices from PDFs and images.– Assisted in development of a pipeline to capture and recognize license plates from video streams, removing the need of manual monitoring.