I was studying engineering when I watched one of Neel Nanda’s videos on mechanistic interpretability, and it moved me enough to switch my major to data science.
The questions I keep circling: how do we decode activations into language we can actually trust, when today’s methods hallucinate or read out little more than a bag of words? How do we get from knowing what a model represents to showing that representation caused what it did? What does it take to probe for something a model never says out loud — deception, or the sense that it’s being watched?
I’m currently working with Dr. Lu Lin at Penn State’s College of IST on interpretability research. I’ve completed BlueDot Impact’s Technical AI Safety course, and start their Technical AI Safety Project Sprint in October.
I’ll be at COLM in San Francisco this October — feel free to reach out if you’ll be there too.
Apart from research, I like music, chess, poker, pickleball and badminton, or just Xbox with friends.
Co-first author · ICML 2026 Mechanistic Interpretability Workshop, Seoul · COLM 2026 Actionable Interpretability Workshop
Showed that SAE decoder-space geometry predicts how steerable a feature is before any intervention, tested across Gemma-2, Llama-3.1, and Qwen3. Worked under Girish Gupta and Aditya Shah, through Algoverse.
paper··
ConfAIde · Under review
Audited a probe that seemed to predict, from Qwen2.5-7B-Instruct’s activations, whether it would refuse to reveal a secret in ConfAIde role-play scenarios. Re-annotating the responses showed the “refused” label mostly marked answers that disclosed the secret and then hedged, so the probe was reading the labels, not the model. What survives is narrower: whether the model will limit a disclosure it is about to make is decodable before the first response token, with no comparable signal in Llama-3.1-8B-Instruct.
·
Mechanistic KYC · Ongoing since May 2026
Probing a frozen Gemma-2-9B financial advisor to see whether it internally represents client risk tolerance differently from what it outputs. Built a synthetic client-profile generator grounded in the Grable & Lytton risk scale.
·
