Revamped Domain-based Relative Positional Encoding for LLMs:Developed a novel relative positional encoding method that uses a domain-specific embedding layer to incorporate domain expertise into the encoding process.Conducted an In-depth Comparison of BERT and LSTM:Using data augmentation for the smaller dataset, a BERT model from Hugging Face Transformers was trained on a small dataset of reviews of tourist attractionns.Outperformed LSTM models trained on larger datasets by 6% using the Ray Tune hyperparameter optimization framework to discover the optimum hyperparameters for the BERT model.Created Tokenization Stability Index:Created a cutting-edge metric to gauge the stability of tokenization algorithms. The performance of tokenization techniques(BPE, WPC, UNG, Sentence piece) is assessed and compared using the measure.Created a RESTful API in Flask that compares documents and determines the similarity score for the use of redundant checking and reduces dataset size by up to 10%.Sentiment Analysis:Developed a BERT sentiment analysis algorithm with the potential to help businesses understand client demands, enhance products and services, and boost salesCompared to the state-of-the-art results, the model outperformed them by 2%, achieving a 95% accuracy on the GLUE benchmark.Developed a Restful API for Similarity Analysis:Designed and implemented a robust RESTful API using Flask to assess document similarity reducing a 30% reduction in document assessment time.Conducted comprehensive text analysis to provide detailed insights into document similarities and differences.Utilized a sophisticated NLP pipeline that included text pre-processing, tokenization, named entity recognition, and dependency parsing.

Empowered Information Retrieval:Engineered & deployed IR system combining TF-IDF, BM25, language modeling, boosting e-commerce catalog searchability.Advancing Clinical Decision Support LLM: Created a named entity recognition (NER) model using spaCy to identify the medication names in the ICD-10 dataset.Trained a Llama model on the NER-processed ICD-10 dataset to predict disease diagnoses based on prescriptions.Demonstrated the potential of LLMs for this task by having the Llama model outperform other methods on a held-out test set through the integration of NER.Soft Prompting LLM for Clinical Decision Support:Customized tokenizer and created NER for Llama for improved clinical language processing and responsesDeveloped adaptive soft prompts backing up continuous, context-aware suggestions in clinical recommendationsIntegration of clinical prompts based on a clinical workflow with continuous recalibration based on real-world feedback

Text Model OptimizationCreated ready-to-use Text Corpus for low-resource Dravidian languages in researching deep learning models.Worked with the AI development team to transfer research work into production.Taught graduate courses on High-Performance Computing, Software Modelling and Design, Machine Learning with 100% practical examples.