My current work starts from one thesis: AI infrastructure is moving from serving models to coordinating intelligence.
The important question is no longer only which model is strongest. It is how to route semantic work across models, tools, memory, verifiers, hardware, and energy budgets so the whole system behaves intelligently.
- Intelligent routing turns workload signals into capability paths across models, tools, memory, and policies.
- Agentic intelligence studies how agents use feedback, verification, memory, and tools to keep improving inside long-running loops.
- System co-design connects algorithms, serving systems, policies, and hardware constraints instead of optimizing them in isolation.
The thread connecting these directions is system-level design: how routing, agents, models, and infrastructure can be co-designed instead of treated as separate layers.