I'm a researcher and founder working on open-ended scientific discovery. I run Dynamical Systems, where I build environments, evaluations, and verification systems that make scientific work trainable.
My work sits across RL, post-training, agent evaluation, and scientific ML, turning search, uncertainty, revision, tool use, and verification into learning signals for models operating in long-horizon environments.
Verified campaign environments convert search, trust, escalation, and revision into a multi-turn RL problem with physics-grounded oracle reward.
Scaling test-time verification for novel materials
Probe-gradient guidance extracts band-gap signal from an unconditional crystal diffusion model and steers sampling without retraining.
Self-improving pretraining as a substrate for agentic post-training
Synthetic thinking traces and self-improvement loops as a substrate for training models that can revise, critique, and extend their own work.
ATLAS: Adaptive Test-Time Learning for Agentic Systems
A continual learning framework that converts production agent trajectories into inference-time adaptation and on-policy distillation loops.
I contribute to inference and training infrastructure in the open-source ML stack:
Primary stack: Python, Rust, PyTorch, Ray, SGLang.