Openings
I am currently seeking motivated students to work on the projects listed below. Each project is accompanied by a GitHub repository that interested students are expected to reproduce before applying and subsequently prepare a report summarising their findings.
Apply by submitting the Google Form, uploading the report, CV, marksheets of all previous semesters, and current CGPA.
Shortlisted students will be invited for a discussion. Selection will be based on your understanding of the project as demonstrated in the reproduction, subsequent discussion, your academic performance, and work ethic. I look forward to hearing from you.
Description: This project develops retrieval-augmented question answering systems grounded in Indian statutes and case law. Students will work on legal document retrieval, citation-grounded answer generation, and generation of hallucination-free outputs, using benchmarks built for the Indian legal domain.
Prerequisites: Proficiency in Python; familiarity with NLP fundamentals and LLM APIs or Hugging Face; interest in law is desirable. Applicants must reproduce this Github link.
Machine Translation and Evaluation for Indian Languages using Agentic AI
Description: This project aims to build a multi-agent framework to automatically evaluate and categorize error types for translation of Indian languages.
Prerequisites: Proficiency in Python and PyTorch; familiarity with agentic AI, sequence-to-sequence models, and knowledge of an Indian language beyond English is an advantage. Applicants must reproduce this Github link.
In-Context Learning in Large Language Models: An Information Retrieval Perspective
Description: The performance of in-context learning (ICL) in large language models is highly sensitive to how demonstration examples are selected, ordered, and combined in the prompt. This project studies these choices from an information retrieval perspective. Students will investigate retrieval strategies, learning-to-rank approaches for example ordering, diversity-aware selection, and adaptive choice of the number of examples, and analyse their effect on performance and stability across tasks and models. The project lies at the intersection of NLP and information retrieval.
Prerequisites: Familiarity with Machine Learning, Deep Learning, or Natural Language Processing. Proficiency in Python and PyTorch; familiarity with information retrieval is an advantage. Repository: https://github.com/janak11111/Diverse-ICL
Machine Learning Modelling of the DHFR Folate Metabolic Pathway and Mutations
Description: Dihydrofolate reductase (DHFR) is a key enzyme in the folate metabolic pathway and a target in drug resistance studies. This project applies deep learning and reinforcement learning to computational biology problems centred on DHFR. Directions include graph-based modelling of metabolite-protein interactions in the folate pathway, and reinforcement learning for exploring and optimising protein sequence variants to study the effect of mutations. No prior background in biology is required, but students are expected to learn the relevant biochemistry.
Prerequisites: Sound knowledge of Deep Learning or Reinforcement Learning. Proficiency in Python and PyTorch.
Repository (deep learning track): https://github.com/mmetalab/mpi-vgae
Repository (reinforcement learning track): https://github.com/haewonc/LatProtRL
Semantically Aware Role-Based Access Control
Description: Role-Based Access Control (RBAC) is the dominant access control model across commercial, government, healthcare, finance, and education sectors. It introduces roles as an abstraction between users and permissions, so that changes in responsibilities require only updates to role assignments. Traditional role mining extracts roles from user-permission matrices using purely combinatorial patterns. This project develops semantic role mining methods that incorporate domain knowledge, learned permission embeddings, and natural language processing, so that the resulting roles align with business functions and organisational structure.
Prerequisites: Sound knowledge of Machine Learning. Experience using LLMs; familiarity with database systems.
Reading: https://dl.acm.org/doi/10.1145/2871148
Generative AI and Design for Traditional Idol-Makers
Description: This project builds and evaluates generative AI tools for the traditional idol-makers of Kumartuli, Kolkata, as part of the ShilpiBondhu application. Students will work on artisan-grounded image generation, region-selective image editing and prompt repair, classification of craft attributes from annotated data, and evaluation of the tools with artisans. The project combines generative AI with human-computer interaction and design research.
Prerequisites: Familiarity with at least one course in Machine Learning, Deep Learning, or Natural Language Processing; and experience in app development. Familiarity with diffusion or vision-language models is an advantage. Knowledge of Bengali is an advantage.