Research

I did my PhD in computational plasma physics at the University of Maryland, then left active research for industry. A decade later, AI has made it possible to come back — not as a full-time researcher, but as someone who can run real physics investigations as a side project.

Plasma turbulence, with AI

Over a couple of months, working largely solo with Claude, I rebuilt my PhD-era gyrokinetics code in JAX and took it all the way to a publishable result on the dissipative anomaly in plasma turbulence — a numerical test of Onsager's conjecture in a kinetic-reduced MHD (KRMHD) system.

The interesting part wasn't that AI could write the code. It was that AI could do the unglamorous, undergraduate-level research motions — cluster runs, config iteration when the physics looked off, diagnostic plots, numerical debugging, paper drafting — while I provided the physics framing, the validation, and the editorial discipline. The bottleneck turned out to be a systems problem (persistent memory, git discipline, cheap decision-making, cross-tool orchestration), not a model-capability problem.

• Paper: github.com/anjor/gandalf-paper
• Code: github.com/anjor/gandalf

Write-ups

• AI as the Undergrad Researcher — the full account: a real physics result, two months, one person
• Building a Gyrokinetics Code Without Reading a Single Line — rebuilding the simulation in JAX with Claude in 30 days
• The Autonomy Gradient — what AI can and can't do across research tasks
• Physics-Oracle Validation — how to trust code you've never read
• Writing a Physics Paper with Claude — workflow, iterations, and hallucinations caught in review