I build LLM systems and developer tools with an emphasis on reproducible experiments, measurable performance, and clear engineering trade-offs.
Computer Science & Engineering at Ewha Womans University.
Extended verl with phase-aware model residency, CPU AdamW, and bounded
asynchronous gradient streaming for memory-efficient full-parameter GRPO.
- Reduced the Actor-backward peak by 27.9% and end-to-end step time by 12.1% with a 16 MiB streaming pipeline.
- Enabled Qwen2.5-1.5B training on a GPU with 11.90 GiB usable capacity where all-GPU and CPU-Adam-without-streaming configurations reached OOM.
- Preserved experiment configurations, raw evidence, failure cases, and known limitations for reproducibility.
Built a reproducible benchmark that separates the effects of context growth, prefix stability, output length, and tool latency on LLM response time.
- Measured TTFT growth from 0.084 s to 0.458 s as controlled context grew from 90 to 24,717 prompt tokens.
- Combined client-side latency with vLLM Prometheus metrics to distinguish KV recomputation, cache reuse, prefill, and decode costs.
- Added a versioned experiment spec, CLI, unit tests, and automated CI checks.
- LLM inference and training systems
- Developer tooling and reproducible evaluation
- Performance measurement and resource-efficient systems
- Interfaces that are understandable to both developers and LLMs
