hey, i'm Swastik

My current research spans mechanistic interpretability and AI-generated text detection. In Dr. Sarah Wiegreffe's group at UMD, I'm investigating the performance gap in activation steering between CAA-style behaviors and SAE-derived concepts, carefully controlling for model, layer, dataset generation, and LM judge. I've also been working on human vs. LLM evaluation comparative analysis to improve judge pipelines for more reliable steerability assessment.

Through the Supervised Program for Alignment Research (SPAR), I'm developing Circuit Oracle — an agentic pipeline for automated end-to-end circuit discovery, interpretation, and causal hypothesis generation in LLMs. This involves conducting jailbreak circuit discovery on Qwen-3-4B and Gemma-2-4b-it against GCG-optimized attack prompts.

In Dr. Fardina Alam's group, I'm co-leading research on training-free AI-generated text detection using signals from letter distributions. We introduced LD-Score, a complementary detection signal that improves AUROC, F1, and accuracy when integrated with Binoculars and DNA-DetectLLM. We also extended the HC3 benchmark with additional models, domains, and temperature settings.

Previously, I spent over a year in the Risk-Informed Solutions in Engineering Laboratory where I engineered a temporal deep-learning model to impute unavailable data on tropical cyclones, improving existing shallow approaches by 15%. That work was presented at ICOSSAR 2025, the AGU Fall Meeting 2024, and the UMD CEE Symposium 2024. Read the paper.

I've also worked in the Tubaldi Lab, designing a multimodal object classification model integrating pressure sensing and visual input for a soft robotic gripper arm, and developing communication frameworks for the UR3e robotic arm with automated trial randomization and data collection.