LLM inference, serving, and systems that should survive inspection.
I'm Jooho Lee (BWAAEEEK), an AI engineer and researcher focused on LLM inference performance, serving reliability, on-premise deployment, and RAG systems.
- My systems work spans Triton kernels, quantized inference, CUDA Graphs, speculative decoding, multimodal streaming, and repository-level context generation.
- I work as an AI / LLM systems engineer, building enterprise on-premise AI products while also working across graph learning and computational pathology.
- I prefer systems that start from a clear problem definition and can be inspected, measured, and improved over time.
- vLLM — Triton sparse GELU, CuTe-DSL MXFP4, speculative decoding scheduling, Qwen3-ASR streaming, audio resampling, encoder CUDA Graph profiling
- LiteLLM — Hugging Face embedding usage
- SGLang — Gemma3 RoPE handling
Python PyTorch Triton vLLM SGLang CUDA Graphs CuTe-DSL RAG Computational Pathology
- GitHub: @BWAAEEEK