Ideas:
“
My work is shifting toward the theory of composition in agentic systems: which formal objects let bounded reasoning units combine into systems you can analyze rather than only tune. Two directions sit at the center. The first is cognitive kernels, treating a reasoning primitive as a kernel in the analytic sense, so that composition and transport between reasoning states have a precise algebra instead of an ad hoc prompt chain. The second is mixture-of-experts orchestration, where routing and gating are the actual decision problem, and where the multi-agent and MoE views of specialization read as the same routing geometry seen twice. Both feed back into the applied work: LLM coding agents under hard constraints, retrieval versus long context in production RAG, and reinforcement learning for execution in algorithmic trading. The constant is a preference for structure over heuristics, and for reporting where the structure breaks.
”
Feedback