I'm a Master's student in Computer Science at UC San Diego, focused on ML systems, the software and infrastructure that make machine learning models efficient and usable. Lately that means sparse attention for long context LLMs, fused GPU kernels in Triton, and ML scientific tooling as DuckDB UDF extensions.
Before UC San Diego, I served six years in the U.S. Army in EOD and logistics, including leading a platoon that cleared more than 1,000 landmines and unexploded ordnance in South Korea. EOD gave me my first taste of deeply technical, multidimensional problems that reward ingenuity and contain tradeoffs. Software gives me that same challenge at lower stakes. After the Army, I moved into software engineering at Palo Alto Networks, where I led frontend integration of an AI copilot into Strata Cloud Manager. I came to UCSD to understand the systems beneath the software I was building.
Outside of that I like running, adventuring and playing guitar. I can now play Adam's Song by Blink-182.
PROJECTS
- Co-authored Double-P, a hierarchical sparse-attention method for long-context LLM inference; analyzed compute-matched top-k budgets and found they violated the top-p attention-mass criterion for over 20% of evaluated heads, motivating adaptive token budgets.
- Implemented C++ and Rust zero-copy memory transfer pipelines via the Apache Arrow C Data Interface (ArrowArray / ArrowSchema), marshaling sparse COO matrices between DuckDB vector storage and Rust ML runtimes.
- Implemented memory-bounded in-database TreeSHAP attribution streaming, dynamically deriving batch sizes to enforce a < 80 MB memory ceiling (8 bytes / attribution) against 200k+ dimensional features without OOMs, supporting browser WASM.
- Optimized sparse matrix scanning across 13M+ cells using DuckDB's UnifiedVectorFormat and flat memory layouts (ShapValuesFlat), eliminating intermediate heap allocations and unnecessary materializations.
- Implemented and benchmarked a fused FP16 matrix multiplication, elementwise add, and ReLU kernel in Triton. Evaluated five tile configurations and achieved a 1.28× speedup over the PyTorch reference (0.265 ms vs. 0.339 ms) on an NVIDIA A10 for a 2048×2048×2048 workload.
- Reimplemented and pretrained a BiomeGPT-style transformer. Designed a Captum Integrated Gradients baseline (species identity + 0 abundance) that resulted in a near-0 completeness delta compared to the all-zero embedding baseline.
- Showed that [CLS] attention cannot reveal taxon attribution polarity with respect to the model's classification decision.
- Implemented cases in Loma's forward and reverse mode autodiff passes, extending existing intermediate representation (IR) mutators to transform program IR into first-order derivative compiled C functions.