• Prepared real patient conversation data for LLM finetuning, optimizing dialogue systems following the Hugging Face chat template.• Working on optimizing the Mistral model for conversational AI, using bitsandbytes, 4-bit quantization, QLoRA, along with Langchain, Hugging Face, SFTTrainer to produce automated chit-chat sessions between real patient and AI, simulating authentic patient-therapist conversations. Achieved a token generation rate of 70 tokens/s on Nvidia T4/a10 GPUs through open-source frameworks. Implemented an inference using ExLlamaV2.• Developed a BiLSTM-CRF model from scratch for NER with 181 entity tags and an F1 score of 85%.• Implemented an intent classification model using PyTorch custom dataset and BERT fine tuning achieving over 90% F1 score and accuracy.

• Established a comprehensive database for high-value non-profit foundations and grantees, implemented an end-to-end Ingest to Process pipeline that incorporated amended tax filing, and generated normalized dataframes for feature engineering.• Led the fine-tuning and refinement of BERT, RoBERTa, BART Zero-Shot text classification, and semantic text matching using SimCSE embedding for sequence classification on the philanthropic data of the foundation grantee network of the entire US non-profit sector to develop a custom-trained model. Also demonstrating expertise in large-scale data processing by implementing the unsupervised LLMs using batches in IU high-performance computing environment.• Developed a prompt engineering chain of thoughts model to capture missing Foundation and Grantee mission statements and textual contents using Falcon 7B, Mistral 7B, Neural-Chat 7B-v1, and Zephyr models.

• Predicted the Remaining Useful Life (RUL) of assets for a leading car manufacturing industry by pre-processing sensor data, utilizing Autoencoders for data refinement, interpolating and applying Linear Regression to forecast the health index (HI) until threshold, and using LSTM with time-frequency transformation to detect break points and predict asset failure.• Created an 80% efficient model in NLP using BERT and Transformers for a topic modelling use case of customer review data.