Raj S. Shah

I’m Raj, a PhD student in Interactive Computing at Georgia Tech (advised by Sashank Varma) and a visiting researcher in Diyi Yang’s SALT Lab at Stanford. My job-talk story is about building AI systems that stay trustworthy when the stakes are human well-being.

I pull on three research threads:

Reliable model operations. Dynamic unlearning evaluations, watermarking, and continual learning recipes that keep LLMs accountable once they leave the lab.
LLMs as cognitive probes. Developmentally staged corpora (BabyLM) and psych-aligned batteries that reveal when models follow-or ignore-human reasoning trajectories.
AI for mental health. Counselor copilots, MI-aware feedback loops, and evaluation protocols that center safety for both helpers and seekers.

Across these threads I ship practical tooling: pip packages for unlearning stress tests, open MI datasets + coaching sandboxes, and benchmark suites for finance, healthcare, and visualization literacy. If you’re evaluating LLMs in the wild (or just want to see the job-talk deck), reach out.

news

Feb 03, 2026	Announced my job talk tour (Georgia Tech → Stanford → CMU HCI)**-demoing our Naïve Scientific Misconceptions probes that uncover where GPT-4o slips back into child theories.
Dec 15, 2025	Started the Stanford SALT Lab residency with Diyi Yang-splitting time between counselor copilots and clarification-driven summarization for Amazon Rufus.
Aug 20, 2025	Received the Georgia Tech President’s Fellowship-three years of support to push on unlearning, mental-health evaluation, and BabyLM.

latest posts

Feb 20, 2026	Domain-Specific Evaluations With Real Consequences
Feb 20, 2026	Teaching Language Models to Grow Up
Feb 20, 2026	Beyond the Unlearning Mirage

selected publications

Helping the Helper: Supporting Peer Counselors via AI-Empowered Practice and Feedback

Shang-Ling Hsu, Raj Sanjay Shah, Prathik Senthil, and 4 more authors

Proceedings of the ACM on Human-Computer Interaction, 2023

arXiv Bib PDF

@article{hsu2023helping,
  title = {Helping the Helper: Supporting Peer Counselors via AI-Empowered Practice and Feedback},
  author = {Hsu, Shang-Ling and Shah, Raj Sanjay and Senthil, Prathik and Ashktorab, Zahra and Dugan, Casey and Geyer, Werner and Yang, Diyi},
  journal = {Proceedings of the ACM on Human-Computer Interaction},
  volume = {7},
  number = {2},
  pages = {1--45},
  year = {2023},
  publisher = {ACM New York, NY, USA},
}

The Unlearning Mirage: A Dynamic Framework for Evaluating LLM Unlearning

Raj Sanjay Shah, Sashank Varma, and others

Proceedings of Machine Learning (COLM) Workshop, 2025

arXiv Bib PDF

@article{shah2025unlearningmirage,
  title = {The Unlearning Mirage: A Dynamic Framework for Evaluating LLM Unlearning},
  author = {Shah, Raj Sanjay and Varma, Sashank and others},
  journal = {Proceedings of Machine Learning (COLM) Workshop},
  year = {2025},
}