At MITRE, I developed human-in-the-loop, tool-augmented LLM workflows to improve user control, interpretability, and reliability. I implemented constrained-output generation that reduced tool-call format hallucinations from >90% to 0%, and presented the results at internal conferences.
I led a project developing a new version of an internal tool used by 1000+ users, integrating agentic hierarchical RAG workflows with human fact verification and feedback. I presented demos, incorporated stakeholder feedback, and secured future funding.
I also architected an RL training system for cyber agents, including Caldera and CyberLayer, and served as an RL subject matter expert for cyber defense via deception research later presented at Black Hat. At Heron Systems Inc. (A Shield AI Company), I used reinforcement learning to train autonomous fighter pilots for beyond-visual-range combat for the Air Force Research Laboratory.
