Hi, I'm Minjun. Most of my time goes into making LLM serving faster and more reliable. I fix bugs and performance issues in open-source projects like LMCache and AIBrix, and I'm also building pygrpc-native, a Python gRPC server backed by Rust. On the side, I make things with friends and write about them on my blog.

LMCache/LMCache · KV cache layer for vLLM/SGLang · 12 merged PRs

• Added a native gRPC server mode in Rust, outside the GIL, with 1.2x ZMQ's throughput (#6031)
• Raised gRPC throughput 1.9x via an asyncio server with bounded stream read-ahead (#5799, #5858)
• Cut gRPC p99 latency from ~9 ms to 1.8 ms by fixing a metrics flush stall (#5760)
• Cut gRPC recovery after a server restart from 36 s to 0.4 s by bounding reconnect backoff (#5825)
• Sped up the gRPC codec for up to 2x lookup and 1.7x stream throughput (#5935, #5906)

vllm-project/aibrix · vLLM's Kubernetes control plane for LLM serving · 11 merged PRs

• Built an encode stage for multimodal requests in PD and E+PD disaggregated routing (#3069, #3153, #3200)
• Cut prefix-cache routing latency by 42% and fixed tokenizer pool failures (#3035)
• Fixed LoRA prefix-cache routing by indexing KV cache blocks per adapter (#3097)

vllm-project/semantic-router · vLLM's intelligent router for mixture-of-models serving

• Cut JSON validation time by up to 44% for requests and 64% for responses (#5204)

vllm-project/guidellm · Benchmarking tool for LLM inference deployments · 2 merged PRs

• Added collectors that scrape server metrics from Prometheus while a benchmark runs (#1250)
• Fixed streaming requests bypassing the mock server's concurrency limit (#1262)

vllm-project/production-stack · vLLM's reference stack for Kubernetes deployment

• Kept prefill/decode model labels when a dynamic config reload rebuilds the router (#1129)

kai-scheduler/KAI-Scheduler · NVIDIA's Kubernetes GPU scheduler · 5 merged PRs

• Cut reclaim time by up to 48% and allocations by up to 75% for many small jobs (#2420)
• Cut memory use by 53% in reclaim scenario search by reusing cloned victim jobs (#2400)
• Cut allocations in GPU node scoring by up to 23% (#2410)

• Software (Computer Science & Engineering), Chung-Ang University (2021.03. ~ )

• Nexters 29th BE Member (2026.07. ~ 2026.08.)
• GDGoC CAU 4th Core Member, 5th Organizer, 6th Team Member (2024.09. ~ )
• Woowacourse 7th BE (2025.02. ~ 2025.11.)
• CECOM Member (2024.09. ~ 2026.02.)
• Undergraduate researcher at HCSLAB of Chung-Ang Univ. (2024.09. ~ 2025.01.)
• GDSC CAU 1st Member (2021.06. ~ 2022.07.)
• CAU ZeroPage 31st Member, 32nd Vice-President (2021.03. ~ )

• pygrpc-native - A Rust-backed gRPC server for Python, used by LMCache. (2026.10. ~ )
• GAMSS - An emotional trash can for things you can't tell anyone. (2026.07. ~ )
• ZZOL - A luck-based game that even clumsy players can enjoy! (2025.07. ~ )