Tathagata Raha

- Developed and integrated a state-of-the-art NER model for identifying clinical entities in texts, enabling conversion of user prompts into SQL scripts for OMOP database queries. - Worked in the development of the Clinical LLM (Med 42). Specifically, setup the finetuning datesets and model evaluation pipeline. Explored different prompt engineering techniques (including few-shot, chain-of-thought, and self-consistency) for enhancing the performance of the models. - Collaborated in the development of a RAG (retrieval augmented generation) system, in which different knowledge bases are used to ground the large language models by providing context information and enhancing the model’s response.

Currently working on a project to point out semantic mismatches in text and explain them Previously worked on depression detection and modeling in social media as a part of Project Angel, explainable fake news detection, and detection of hate speech in a multilingual setting. Published multiple papers in reputed workshops and conferences. Can be viewed in my Google Scholar page.

Role includes but is not limited to: teaching Data Science and Machine Learning/Deep Learning concepts, solving doubts in programming questions, taking mock interview sessions, and evaluating assignments.

Implemented an Unsupervised Document Grounded Dialogue Bot to answer queries after retrieving relevant knowledge sentences. Designed a graph-based approach to capture various inter-dependencies between dialogue utterances and knowledge sentences, model the flow of the conversation and account for the structure of the document while selecting relevant knowledge sentences. Ran experiments with BERT embeddings similarity and TF-IDF-based edge weights in the graph and trained Node2Vec-based node embeddings to achieve a median rank 7 for selecting relevant knowledge sentences.

Worked on building a web app for day-to-day management of Child Development Centres. Tech Stack: Used ReactJS for frontend, Django for backend, and MongoDB for database management.

Developed a POS-tagger for English-Bengali code-mixed data