• Orchestrated the development and deployment of cutting-edge Generative AIapplications for prominent clients across the Insurance and Consumer Packaged Goods(CPG) domains.• Architected end-to-end Retrieval Augmented Generation (RAG) systems,encompassing efficient data ingestion pipelines, vector database integration, semanticsearch, and response generation using Large Language Models (LLMs).• Implemented agent-based system framework using OpenAI Functions and LlamaIndexto handle the correct flow of conversation in a chatbot.• Proficient in fine-tuning open-source LLMs (e.g., GPT-3) for task-specific purposes usingtechniques like LoRA and QLoRA, achieving superior performance and reducingcomputational requirements.• Implemented advanced quantization strategies (GPTQ, AWQ) and optimizeddeployment frameworks (TGI, vLLM, TensorRT, Triton) to minimize infrastructurecosts without compromising model quality.• Established robust evaluation and monitoring methodologies for Generative AIapplications, leveraging tools like MLflow, OpenTelemetry, and Arize AI, ensuringoptimal performance and scalability.• Applied state-of-the-art prompting strategies, including few-shot learning, ReActPrompting, Chain of Thought, Chain of Verification, and Retrieval AugmentedGeneration, to enhance model outputs and accuracy.• Designed and maintained vector databases and semantic search pipelines,incorporating re-ranking layers for efficient retrieval and contextualized responses.• Spearheaded the creation of Point of Views (PoVs) for Generative AI applications,driving client acquisition and fostering innovation.• Leveraged diverse cloud computing platforms, including Azure, AWS, and GCP, todevelop and deploy Generative AI solutions, utilizing cutting-edge tools and serviceslike GCP Vertex AI, Azure Document Intelligence, Azure AI Studio, AWS Textract, AWSKendra, AWS EC2, AWS S3, AWS Dynamo DB and AWS Lambdas.

• IQ Question Answering System: o Developed a question answering system that generated MCQs and fill-in blanks used as a product in Impelsys IQ service. o The model is trained on the SQuAD to multitask dataset and the distractors are created from sense2vec models. o The model was evaluated at an F1 score of 86.67 and an exact match score of 87%. o The application was built both for the general and medical domains. • MLOps using Amazon Web Services: o Developed best practices for MLOps in AWS environment for core IQ product. o Designed architecture in AWS using AWS Sagemaker, AWS Lambda, and S3 which replicated the traditional IQ Architecture on the AWS platform using MLOps. o This resulted in rapid innovation through robust Machine Learning lifecycle management and helped in the creation of reproducible workflow and models, leading to the easy deployment of high precision models at any location, effective Machine Learning resource management and reduced training cost by around 62% or ~$76k, due to auto training feature via Sagemaker.• PDF Preprocessing Pipeline: o Developed a pipeline and its components to extract and preprocess text data from pdfs for IQ services using Grobid client for pdf extraction and standard text preprocessing and cleaning methods. o Implemented Natural Language Processing techniques for text preprocessing such as stop-wordremoval, stemming, lemmatization, Bag of Words, TF-IDF, Word2Vec, and word embeddings.• Part of the core team member of Tech Tuesday and delivered training sessions on Machine/Deep learning topics.

• Evaluated performance and policies against metrics.• Developed and updated tracking spreadsheets using PowerBI and MS Excel.• Developed innovative planning models, reducing lead times 60%.

• Built BERT Hindi language Model on huge Hindi text corpus.• Developed Emotion detection classifier in Hindi using fastbert.