MS in Artificial Intelligence at Columbia University and NSF Graduate Research Fellow. I work on interpretability, vision-language models and AI safety for LLM agents. Before this I did computational neuroscience research at UPRRP, Yale's Wu Tsai Institute and UChicago.
Research
- What Looks Like a Capability Limit in Vision-Language Models Is a Readout Limit (arXiv, 2026). Code, per-trial logs and pre-registration: vlm-readout-limit
- graftpp-collusion-monitoring: team project from the Apart AI Collusion Research Sprint 2026. Monitors change what LLM seller agents say more than what they charge.
Projects
- agent-risk-rails: a pre-execution gate for agents that can move real money
- tiered-classifier: rules first, LLM second, and a guard that refuses invented labels
- canary: team hackathon build that watches a startup's bank ledger for sustained shifts in spending
- app-studio: build and ship iOS apps from Windows with Expo and EAS