Developed LLM-based diagnostic tools (GPT, Meditron, Llama) for veterinary medicine, achieving a 95% validation accuracy across 10,000+ cases using advanced prompting techniques.Built and deployed a bilingual chatbot for HCMC’s Transplant Center, serving over 5,000 patients, integrated with UTMB's website via Azure services, improving response times by 35%.Enhanced health inspector training with a RAG system, increasing training efficiency by 60% and reducing training time by 40% through scenario simulations using SFT and RLHF data.Leading the PASCAL project, applying LLMs to assist 1,000+ patients in quitting smoking, integrating with a custom-built website to track progress and offer personalized support.

Developed a custom Sanskrit language model pipeline by fine-tuning state-of-the-art LLMs (Llama-3.1, Gemma-2) for tasks such as translation and question-answering, achieving significant improvements in BLEU and ROUGE scores across 11,000+ synthetic and real data pairs.Engineered a specialized tokenization scheme for Sanskrit, utilizing Byte Pair Encoding (BPE) and one-hot encoding to handle the complex morphological structure of the language, optimizing vocabulary size and processing efficiency for large-scale language modeling.

Developed a Retrieval-Augmented Generation (RAG) pipeline for qualitative data in a trading application, improving query processing efficiency by 50% through the integration of advanced agentic systems like AutoGen, LangChain, CrewAI, and LlamaIndex to facilitate multi-agent query routing.Engineered RAG Fusion-based query processing, incorporating on-the-fly semantic chunking of SEC Form 10-K and 10-Q financial documents, processing over 100,000 chunks with embeddings stored in vector databases (FAISS, Pinecone, pgVector). Achieved 80% accuracy in document retrieval using cosine similarity, reranking results, and augmenting prompts with the most relevant data.Contributed to the implementation of GraphRAG, increasing retrieval accuracy by 40% by leveraging Neo4j and Cypher queries to build a more robust knowledge graph for financial data, improving decision-making insights.Advanced the state-of-the-art in RAG systems by integrating multimodal data processing capabilities, handling over 50,000 textual and visual data points to improve chunking and retrieval performance in trading applications by 30%.

Utilized quantized large language models to enhance Time Series Forecasting project, resulted in a 20% reductionin forecasting errors and improved overall project efficiency, used SOTA Mamba, TimesNet, Autoformer, NBEATSarchitectures and Prophet library to achieve enhanced performance

Worked on building NLP models for sentiment analysis an Azure Cloud Pipelines.