ABOUT OREN
I'm a Master's student in Computer Science at UC San Diego, focused on ML systems, the layer where models meet hardware. Lately that means sparse attention for long context LLMs, fused GPU kernels in Triton, and bringing in ML algorithms into duckdb database via zero copy apache arrow columnar format.
The road here wasn't a straight line. After a finance degree at Florida State, I commisioned into the U.S. Army as a second lieutenant and spent six years doing various Explosive Ordnance Disposal (EOD) and logistics missions, and led a platoon that cleared 1000+ landmines and unexploded ordnance in South Korea. EOD is deeply technical. Every problem has a dozen dimensions, and it rewards ingenuity and novel solutions. The feedback loop is less forgiving though. ISTF (initial success or total failure).
Leaving the Army, I honestly didn't know what I wanted to do with my life. Then I remembered I used to code in high school on a TI-83+. Long story short, I found myself in a coding bootcamp (one of the last to ride that wave), then in the valley at Palo Alto Networks, where I led the frontend integration of an AI copilot into Strata Cloud Manager. But frontend didn't satisfy my hunger to know what was happening underneath. So here I am at UCSD, going deeper. Turns out software scratches the same itch as EOD, just with much friendlier consequences for being wrong.
Outside of that I enjoy running and adventuring over air, land and sea. I also picked up guitar, and can now play Adam's Song by Blink-182.
A few things I've learned so far:
- The last goodbye is not always obvious. Make beautiful memories with the ones you love. They will remain in the end.
- The Five Remembrances by Thich Nhat Hanh.
- Trust your gut. Listen early so you don't have to repair late.
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.