Systems infrastructure × AI inference × Agent infrastructure.
I work at the intersection of low-level systems and large-scale AI. My background is in distributed databases, storage engines, and disaggregated memory systems. Currently, I'm focused on making LLM inference faster and building the infrastructure layer for AI agents.
CS grad student at USTC. Open-source contributor. Builder.
- Optimizing KV Cache scheduling for large model inference (contributing to Mooncake)
- Exploring video generation / multimodal inference acceleration
- Building infrastructure for AI agents — memory, retrieval, orchestration
- Researching disaggregated memory architectures (CXL, PMEM, RDMA)
Systems Layer AI Infrastructure Frontier
───────────────── ───────────────── ─────────────────
Storage engines LLM inference optimization Agent memory systems
KV stores (LSM / B-Tree) KV Cache scheduling Retrieval acceleration
Transaction processing Distributed training Multimodal agents
Disaggregated memory Video generation systems Research agents
RDMA / CXL / PMEM Serving infrastructure Agentic workflows
Contributions to production infrastructure used at scale:
| Project | What I Worked On |
|---|---|
| RocksDB | Storage engine performance optimization |
| Mooncake | Transfer engine, KVCache scheduling for LLM serving |
| Apache Kvrocks | Storage engine internals |
| OpenMLDB | Database kernel for ML feature platform |
| ColossalAI | Distributed training framework |
| Kmesh | eBPF-based kernel-native service mesh |
| LightGBM | ML framework contributions |
- Mooncake KVCache — Working on the serving platform behind Kimi. Focused on transfer engine and KV Cache disaggregation for efficient LLM inference.
- LevelDB-BF-Index — Bloom filter index optimization for LSM-tree based storage.
- 3R-Memory-Manager — Custom OS memory management with a novel 3R allocation strategy. National OS competition project.
- Leverage over labor. I prefer building systems that multiply output — automation, good abstractions, AI-augmented workflows. Not a fan of scaling through headcount.
- Research that ships. Theory matters when it changes how systems work in practice. I want to close the gap between papers and production.
- Infrastructure taste. The best infra is invisible. I care about clean interfaces, minimal abstractions, and systems that degrade gracefully.
- Email: 782294150@qq.com
- Site: yanchaomei.github.io
Building systems that think — from storage engines to intelligent agents.