AI infrastructure engineer on Sangfor's Managed Cloud Platform team,
building predictable LLM serving, elastic GPU memory, and runtime protocols.
I work where inference engines, GPU runtimes, and cloud-native control planes meet. My current focus is making shared accelerators behave like dependable infrastructure: correct under pressure, observable in production, and fast on real workloads.
- LLM serving: vLLM and SGLang integration, long-context inference, PD multiplexing, CUDA execution paths, and correctness-first performance analysis.
- Elastic GPU memory: KV cache pooling, CUDA VMM, quota and reclamation policy, multi-instance isolation, MPS, and shared model weights.
- Runtime protocols: building NNRP across Rust, Python, and C# with deterministic transports and explicit lifecycle semantics.
- Cloud native: Kubernetes-oriented placement, observability, recovery, and platform integration for production inference systems.
| Area | Project | What I work on |
|---|---|---|
| Runtime protocol | NagareWorks/nnrp-rs | Native transport and lifecycle runtime |
| SDK | NagareWorks/nnrp-py | Python runtime bindings and release engineering |
| SDK | NagareWorks/nnrp-cs | C# protocol SDK and native transport integration |
| Unity tooling | UnityEasyInject | Lightweight dependency injection for Unity |
Measure the real path. Preserve the invariants. Optimize what remains.