This is my ongoing PhD Thesis, where we are currently exploring the problem of automatic short-answer grading with feedback. The task is to assign a level of correctness (Correct/Partially Correct/Incorrect) to a student's answer given the question and reference answer, along with feedback that points out the errors present in the student's answer and ways to mitigate them.

• Research Concepts : Explainability, Causality, Multimodal Reasoning

• Guide: Prof Bhaskaran Raman and Prof. Pushpak Bhattacharyya, Dept. of CSE, IIT Bombay

SAFE is platform which is a complete classroom management system currently being used by IIT Bombay and other institutes. Using SAFE instructors can conduct quizzes, take attendance and apart from this SAFE is robust and capable of detecting malpractices. Currently SAFE does not have any provisions to grade scanned answer sheets. This projects aims at introducing AI Assisted Grading to SAFE wherein SAFE will take in scanned answer sheets and automatically detect name, roll no and questions regions. Next it will use this information to detect handwritten answer regions and henceforth highlight certain keywords present as rubrics. This is my ongoing MS Thesis.

• Technologies Used : Django, VueJS,  Python(Pytorch, Detectron2, docTR)

• Guide: Prof Bhaskaran Raman and Prof. Parag Chaudhuri, Dept. of CSE, IIT Bombay

Apache Spark is a unified analytics engine to process Big Data.  It has over 150 runtime parameters which need to be tuned to obtain the best possible runtime for a particular application running on a cluster.  Now 150 parameters means its a huge search space to find the optimal set of parameters. Can we build a model which will predict runtime give config parameters and evetually vice versa. However these models will still have a huge training time. This RnD aims to check if system performance metrics such CPU utilization and Memory used are related to these spark parameters or not and eventually use this insignt to try and limit this search space by lmiting the range of each parameter. This led to the development of our own monitoring framework and eventually use it to gain some interesting insights into how Apache Spark works.

• Technologies Used : Python, Shell Scripting, LINUX, Java

• Guide: Prof Varsha Apte (HoD), Dept. of CSE, IIT Bombay

Multi-objective Optimization problems have been in existance for quite sometime and one such problem is the Job Shop Scheduling Problem. This project aimed at developing an algorithm which would automatically find the Pareto optimal solutions whenever the permissible range of an objective in consideration was changed in a multu-objective optimizaiton problem. Together we developed a Java library to measure the effects of such fuzzy performance measures in multi-objective problems. Till date we have only tested on the Job Shop Scheduling Problem but we intend to extend it to other problems as well.

• Collaborators :  Dr. Arkopaul Sarkar and Late Dr. Gürsel Süer, Industrial and Systems Engineering Department, Ohio University

Thymoma is a form of tumour which originates from the epithelial cell of the Thymus gland. If left untreated this can become malignant and eventually become life threatening. Early stage detection of cancer can always save lives. This project took into consideration two regions namely the Hypothalamus and the Hippocampus and tissue specific gene specific data was collected from the GTEx portal.  Next the data was categorized into four groups namely Hypothalamus GHR High group, Hypothalamus GHR Low group, Hippocampus GHR High group Hippocampus GHR Low group.  FDR analysis was performed between the corresponding High and Low groups to determine the upregulated and downregulated genes.

• Guide: Dr. Reetobrata Basu, Ohio University

Data  analysis  plays  an  indispensable  role  for  understanding  various  phenomena.  Clustering  problem is  an  important  unsupervised  learning  problem.   It  exists  in  the  fields  of  image processing,   pattern   recognition,   machine   learning   ,   bioinformatics,   information retrieval  and  so  on.   In  this  paper  we  propose  a  new  evolutionary  based  approach  to clustering  problem.  The  algorithm  is  based  on  how  trees  maintain  their  leaves and has  been  tested  on  Iris,  Wine,  Seeds  and  Vehicle  datasets. Tt can be seen that the results  are  quite  promising.  The  algorithm  can  be  heavily  parallelized  depending on  how  one  is  defining  the  operators  and  can  be  tuned  to  perfection  for  other optimization  problems  as  well. This was my Final Bachaelor's Project.

• Guide: Dr. Uttam K. Roy,  Dept. of Information Technology, Jadavpur University

• Team Members : Arjya Das ,

Writing correct, high-performance, and stable multi-threaded programs entails a careful trade-off among various aspects of the program. While conservative synchronization and coarse-grain locking assist in writing correct and reliable programs quickly, they limit concurrency of the program—execution of threads accessing disjoint data sets protected by the same lock are unnecessarily serialized. Lock-free schemes often optimistically provide concurrent data structure implementations without a critical section or a software wait on a lock variable. The Project focuses on optimizing the starvation count in traditional Dining Philosopher's Problem using the Lock-free approach of Software Transactional Memory.

• Guide: Mr. Utpal Kumar Ray,  Dept. of Information Technology, Jadavpur University

• Team Members : Arjya Das ,