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 covers 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. EOD's job is making sure destructive devices, from landmines up to nuclear weapons, don't do unintended things to unintended people. In South Korea, I led a platoon that cleared more than 1,000 landmines and unexploded ordnance just south of the DMZ. EOD gave me my first taste of deeply technical, multidimensional problems that reward ingenuity and contain tradeoffs. It also gave me a sense of purpose helping people. And it turns out computer science has similar kinds of challenges and rewards. After the Army, I moved into software engineering at Palo Alto Networks. I started on Prisma Access, building features and dashboards and optimizing bottlenecks in page rendering. I then spent 2023 and 2024 helping get Strata Cloud Manager off the ground during its push to bring everything into one shared dashboard. In 2024 I led the frontend integration of its AI copilot. By then the frontend work no longer felt challenging and I missed the sense of purpose I had before. So I came to UC San Diego for the computer science education I'd skipped in my transition from Army to industry. Initially I came for AI and biology, then realized I really like understanding the systems that make everything run. AI is starting to write those systems too, and I think it's worthwhile to understand that process.
In my free time, I like running, paddle boarding and playing electric guitar. I'm currently into punk, nu-metal and grunge.
PROJECTS
- Produced the paper’s motivating analysis (Fig. 2): with top-k fixed at 𝔼[Aₚ], the mean adaptive budget over all decode steps, retained attention mass still fell below the target p in over 20% of heads (averaged over layers and decode steps), motivating the adaptive token budgets behind Double-P’s near-zero accuracy drop and up to 1.3× faster end-to-end decoding.
- Independently built an autonomous Claude agent loop with an Nsight Compute guided evaluation harness that profiles each candidate, checks it against an FP32 reference, and iteratively optimizes Double-P’s k-means Triton kernel.
- Achieved a 1.38× median kernel speedup across 16 Qwen2.5-1.5B key sets on an A100 (1.41× on Triton 3.8), executing 31.9% fewer warp instructions. Controlled ablations showed deferring the argmax reduction outweighed score/index bit packing 2:1 (53 µs vs. 26 µs saved). Packing’s 3% gain costs FP32 score precision: results held a 2⁻¹² normalized score gap tolerance at 2,048 centroids but, as predicted, not at 4,096 or 8,192.
- Traced atomic access patterns through TTGIR/PTX/SASS: the agent’s runtime-stride epilogue cut atomic sectors/request from 16 to 4.12 on Triton 3.6, matching Triton 3.8’s upstream coalescing fix PR #10059. Verified that fix on the original kernel and isolated a remaining runtime gap it does not close.
- Added random-forest classification and regression to DuckDB as 9 SQL functions (fit, predict, predict_proba, TreeSHAP, feature importances, model features, cross-validation), replacing an out-of-process Python/scikit-learn path that could not run in the browser. Validated against scikit-learn-derived reference bands.
- 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 TreeSHAP streaming that derives batch size from an 80 MB attribution budget for memory-constrained browser WASM (2,000 samples per batch at 5,000 features). In Rust, replaced nested Vec<Vec<Vec<f64>>> results with strided writes into one contiguous buffer exported zero-copy to Arrow, cutting peak matrix copies from 3 to 1 and incremental memory from 282 MB to 86–95 MB at equal runtime.
- Eliminated ~5 heap allocations per cell when scanning a 13M-cell, 197K-feature table by reading DuckDB’s UnifiedVectorFormat as std::string_view and interning sample and feature IDs directly into a sparse coordinate (COO) matrix.
- Exposed the model’s training feature vocabulary through sc’s C API so predict aligns columns to it, preventing silently wrong predictions when inference data has a different feature set.
- Benchmarked duplicate-cell detection over 13M cells: packing (row, col) pairs into sorted uint64s beat a hash set by ~2× in speed and ~4× in memory, which matters under WASM’s 4 GiB linear-memory ceiling.
- 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).
- 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 attribute each species-abundance token's contribution to the classification decision, proposing a source-derived baseline (species embedding + MLP(0)) that gave much smaller convergence deltas than a zeros_like baseline in preliminary runs.
- 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.
- Built a distributed video service in Go and gRPC: videos are segmented into MPEG-DASH chunks and placed across storage servers on a consistent-hash ring, so storage scales horizontally.
- Stored metadata in a 5-node etcd cluster (Raft consensus) that stays available with up to 2 nodes offline; an admin service redistributes chunks as storage nodes join or leave the ring.
- 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.
- Won Best Use of Meshy and Best Use of XREAL Glasses at MIT Reality Hack 2024 with a 5-person team.
- Contributed pulseIn() to WiringPi (3.3K★), the C GPIO library for Raspberry Pi: measures the width of a HIGH or LOW pulse in microseconds with a timeout, for reading pulse-width sensors such as ultrasonic rangefinders.
- Merged in 2025 and shipped in v3.20 (Sept 2026); the maintainers later reimplemented it on interrupt events to cut CPU use.