RL-Based Data Attribution for PPO
Designing an RL-based data attribution framework that identifies the training trajectories driving PPO policy decisions in time-series foundation models, improving attribution precision by 20% over baseline methods.
Projects
Systems and ML projects from research labs and courses. Most of the code is on GitHub.
Designing an RL-based data attribution framework that identifies the training trajectories driving PPO policy decisions in time-series foundation models, improving attribution precision by 20% over baseline methods.

A compiler for a C subset that translates 3-address code with a symbol table into GCC-compatible assembly, with lexing in Flex and parsing in Bison. Earned an AS grade, ranking first in the department.
View code
Extended xv6 with POSIX-like kernel threads, spinlocks and mutexes, and a hybrid SJF/Round-Robin scheduler. Optimised memory management with lazy allocation and swap-based paging.
View codeA Unix-style shell written in C from scratch, supporting command parsing, process control (fork/exec/wait) and I/O redirection. Earned an AA grade, ranking first in the department.

A framework to estimate the minimum number of annotators needed for reliable labels in NLP datasets, applied to toxicity classification of social media comments to detect labelling inconsistencies and cut annotation effort.
View code
An immersive 3D VR tour of Majuli, the world's largest river island and the cultural capital of Assam, built to help preserve its cultural heritage.
View code