Hi! I’m Achu. I’m an AI safety researcher and engineer working on model evaluations, mostly for LLM agents and, more recently, for LLMs controlling robots.

I’m currently a founding engineer at Robocurve, where I work on evaluating frontier models on real robot hardware. I’ve written our reports on how GPT-6 Astra and Fable 5.1 handle manipulation tasks, co-authored RoboHarm (which found that models that refuse harmful requests in chat mostly carry them out once they control a robot arm), and I run our open-source benchmarking program. Before that, I was a MATS fellow with David Lindner, studying how agent harnesses affect model safety. Earlier I worked on evaluation awareness with Santiago Aranguri, and on exploration hacking and goal drift in agents.

I’ve had fairly diverse career experiences in the past few years - before getting into AI, I worked as a Data Scientist at Jerry, and prior to that I majored in Molecular Engineering at the University of Chicago after initially intending to pursue Biology.

Outside of research, I enjoy thinking about, watching, and playing football. I love reading (audiobooks) and I like to sing whenever the opportunity arises (in a past life I was really into a cappella).

Feel free to reach out if you’d like to chat :)

Email • CV • Google Scholar • GitHub • LinkedIn