Hi!
I am a Lead Machine Learning Engineer at Citi, building AI-enabled products for finance using large language models (LLMs) and advanced NLP techniques.
I have more than 8 years of work experience in Data Science and Machine Learning. I have collaborated with Wellma Health in developing Generative AI applications using retrieval augmented generation (RAG).
I have done few open-source contributions at GitHub. My current interest lies in LLM fine-tuning, reasoning, alignment of LLMs, and RAG applications.
- Model fine-tuning/alignment: QLoRA, DPO, LoRA, ORPO
- Deep Learning: PyTorch, Tensorflow, HuggingFace
- MLOps: FastAPI, MLFlow, Git, Docker, Kubernetes
- Large Language Models (LLMs): Llama-3, Mistral, BERT, Flan-T5
- AWS: EC2, S3, Lambda, Sagemaker
-
Fine-tuning of open-source LLM using QLoRA
(Technical Blog) Fine-tuned a Falcon-7B large language model on custom mental health conversational dataset using Low-Rank Adaptation of Quantized LLMs (QLoRA). The dataset was curated from online healthcare blogs, FAQs etc and removed personally identifiable information (PII). Achieved a ROUGE-1 score of 0.37 on the test dataset. -
Prompt Testing Framework for LLMs
(GitHub Code) Designed a prompt testing framework for LLMs to compare and test prompt quality for multiple system prompts based on the generated answers. Measured answer accuracy using multiple NLP metrics i.e. ROUGE, BLEU, and BERTScore and multiple Responsible AI metrics i.e. Faithfulness, Answer Relevancy Score, Harmfulness etc. Achieved a 27% improvement in system prompt quality for different LLM applications. -
Youtube AI Assistant using Langchain
(GitHub Code) Designed an end-to-end Retrieval Augmented Generation pipeline to generate summarized text using GPT-3.5-Turbo LLM for YouTube video transcripts and built a conversational AI system to get instant answers for YouTube video transcripts. Used Qdrant Vector DB to retrieve transcript chunk embeddings and GPT-3.5-Turbo LLM from Langchain to generate answers thereby achieving BLEU score of 0.75. -
Product Description Generator
(GitHub Code) Generated SEO-compliant product description using product title and meta-keywords provided by the user. Implemented custom LLMChain function from the Langchain and few-shot prompting technique to generate multi-paragraph rich text product description for each product name and its corresponding keywords. Built a gradio app as a frontend demo to showcase the technique.
