A mix of research codebases, course projects, and things I built to learn. Everything is on GitHub.
Evolutionary pruning framework for large language models. Treats subnetworks as organisms in a population, evolving compact architectures through selection, crossover, and mutation. The codebase behind our AAAI 2026 Student Abstract on "Natural Selection of LLMs via Pruning."
An investigation into whether sycophancy and sandbagging share mechanistic roots inside transformer models. If they're mechanistically distinct, safety researchers need separate classifiers and steering vectors for every failure mode — if they're not, a unified intervention might suffice.
Implementations and experiments around reinforcement learning from human feedback — reward modeling, PPO fine-tuning, DPO, and various optimization techniques for aligning language models with human preferences.
Exploring multi-objective preference optimization for language models — aligning models to satisfy multiple, potentially conflicting, human preference dimensions simultaneously.
A pipelined RISC-V processor implementation in Verilog. Includes the standard 5-stage pipeline with hazard detection, forwarding, and branch prediction. Built as part of computer architecture coursework.
A computer vision pipeline for detecting and classifying facial features. Uses convolutional neural networks for feature extraction and classification across standard face attribute benchmarks.
Course project for the DAC-202 module at IIT Roorkee. Covers foundational data analysis and computational methods.