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 scalable. My recent work has been about recursive self improvement for kernel development and scientific ML tooling built as DuckDB UDF extensions.
Before UC San Diego, I served six years in U.S. Army 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 a feel for that same challenge. 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.
In my free time, I like running, paddle boarding and playing electric guitar. I'm currently into pop-punk, nu-metal and grunge.
PROJECTS
Double-P: Hierarchical Top-P Sparse Attention for Long-Context LLMsPAPER
2026
UCSD Stable Lab
- Analyzed sparse attention with top-k = 𝔼[Aₚ], where Aₚ is the known total adaptive budget over all decode steps. Found mean retained mass across layers and query heads fell below p in over 20% of decode steps, motivating adaptive token budgets.
- Independently developed an autonomous agent loop with Nsight Compute guided evaluation harness to profile and recursively self-improve a k-means DSL kernel from Double-P.
- Achieved 1.38× median kernel speedup across 16 Qwen2.5-1.5B key sets on A100. Independently isolated deferred reductions and atomic accumulation changes through controlled ablations, validating results within a 2⁻¹² normalized score gap tolerance at 2,048 centroids.
- Traced inefficient atomic access patterns through TTGIR/PTX/SASS in Triton 3.6 on an A100 and verified that Triton 3.8’s upstream coalescing fix PR #10059 reduced sectors/request from 16 to 4.12, while identifying a remaining performance gap despite equal sector efficiency
DuckDB-miint: ML integration into a columnar DB extension using Apache Arrow
2026
UCSD Knight Lab
- Implemented zero copy, in-process interoperability between C++ DuckDB UDFs and a Rust random forest library through the Apache Arrow C Data Interface (ArrowArray / ArrowSchema), transferring buffers across the FFI boundary without serialization or data copies.
- Engineered batched, in database TreeSHAP attribution streaming for datasets with 197K sparse features, dynamically sizing batches to keep attribution buffers below an 80 MB budget for memory constrained browser WASM execution. Replaced nested Vec<Vec<Vec<f64>>> results with direct writes into a contiguous buffer using strides and zero copy Arrow export, eliminating two full matrix copies and reducing measured incremental peak mem from 282 MB to 86 MB
- Optimized sparse input scanning across 13M+ cells using DuckDB’s UnifiedVectorFormat and std::string_view, avoiding temporary string materialization before interning sample and feature identifiers into a sparse coordinate (COO) representation.
Triton Fused Kernel and Distributed training primitives
2026
CSE 291: ML Systems
- 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.
- Implemented NCCL style distributed training primitives, binary-tree All-Reduce, point-to-point All to All, via Open MPI, to support data, tensor and expert parallelism (MoE).
Token-Source Attribution for Microbiome ModelPAPER
2026
- Reimplemented and pretrained a BiomeGPT-style transformer and investigated whether [CLS] attention reflects the direction of individual taxa’s contributions to classification decisions.
- Used integrated gradients to understand each species-abundance token's contribution towards the final classification decision. Used a feature tokenized baseline as species embedding + MLP(0) that resulted in a near-zero completeness delta compared to a zeros_like embedding baseline.
ARM SVE DGEMM kernel
2026
CSE-260: Parallel Computing
- Took a BLIS style 5 loop setup to 97% of vendor BLAS (Arm Performance Libraries) GFLOP/s using cache blocked page aligned buffers, broadcast outer GEMM SVE intrinsics, and grid search autotuned params.
- Engineered a custom 4x16 microkernel utilizing 256 bit vector registers; 16 for accumulation, 8 for loading, and 8 for pipelining and predicates. Used non temporal stores to reduce cache pollution.
Loma Differential Programming DSL Compiler
2026
CSE 291: Differential Programming
- 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.
// from back in the days when I used to tinker with AR
Earth
Backyard Office Pod
1:1 digital twin for my cousin's pod business
Pineapple Dude
3D models by others. I converted them to USDZ for iOS and served them with model-viewer.