Senior military officer and operations research scientist applying high-performance computing and stochastic programming to the hardest resource allocation problems in defense.
I am a Lieutenant Colonel in the United States Marine Corps and a fellow in the Marine Corps Doctor of Philosophy Technical Program (PHDP-T) at the Naval Postgraduate School, competitively selected to conduct doctoral research in Operations Research. My work sits at the intersection of GPU computing, mathematical optimization, and defense decision-making.
My dissertation investigates GPU-accelerated Splitting Conic Solver (SCS) as a competitive alternative to classical scenario decomposition for large-scale two-stage stochastic linear programs — with application to a novel stochastic programming formulation of Marine Corps munitions procurement under uncertainty. I have presented this work at the International Conference on Stochastic Programming (Paris, 2025) and the SIAM Northern and Central California Sectional Conference (2025), and will present at SIAM OP26 (Edinburgh, June 2026).
Prior to my current assignment, I served as an Operations Research Analyst at Manpower and Reserve Affairs, HQMC, where I built analytical tools supporting the Marine Corps' $2.3 billion BAH appropriation program and applied optimization methods to officer talent management policy.
Dissertation and active research threads
PhD Dissertation · Expected Summer 2027
Investigates GPU-accelerated Splitting Conic Solver (SCS) as a competitive alternative to classical scenario decomposition (Progressive Hedging) for large-scale two-stage stochastic linear programs. Two contributions: (1) a computational study of GPU-SCS vs. Progressive Hedging and Gurobi on SIPLIB benchmark instances, and (2) a novel two-stage stochastic programming formulation of Marine Corps munitions procurement under uncertainty — a problem class currently addressed by discrete scenario selection methods rather than full stochastic programming.
CUDA kernels written via Numba that exploit the block-angular structure of the two-stage stochastic program coefficient matrix — parallel sparse operations for K, KT, and (I + KTK) across the A, Ts, and Ws sub-matrices. Structure-aware CPU implementation achieves 5–14× speedup over naive baseline; GPU implementation operational with kernel tuning ongoing.
Evaluated Ruiz normalization, adaptive/warped proximal operators, and Anderson acceleration on 2SSP instances (10–1,000 scenarios). Structure-aware solver achieves 14–274× speedup over standard method. Combined enhancements reduce iterations by up to 67%. CG iteration limiting is the dominant source of practical acceleration. 64-bit precision required — 32-bit arithmetic leads to numerical instability.
Novel two-stage stochastic programming formulation of Marine Corps munitions procurement. Current DoD planning methods select from discrete scenarios — this work formulates the full stochastic program and evaluates GPU-SCS on the resulting instances, connecting computational methods directly to an unclassified defense planning application.
Does Ruiz scaling help or hurt convergence on two-stage stochastic programs? Can effective preconditioning enable 32-bit arithmetic on consumer GPUs? What is the crossover point at which GPU-SCS becomes competitive with Progressive Hedging? These are active investigation threads.
Selected technical projects spanning GPU kernel optimization, operations research, and network optimization. Each project reflects hands-on work at the intersection of computing and applied mathematics.
GPU-accelerated computation of the Mandelbrot set in Julia using CUDA.jl. Achieved 119,315 GFLOPS on RTX 4080 via FMA operations, loop unrolling (@unroll 128), early divergence exits, and memory-coalesced access patterns. RTX 4080 outperformed the A100 by 2.81× on this workload — illustrating the importance of memory bandwidth over raw compute for certain GPU kernels.
Network flow model for crisis evacuation routing using a directed graph with 3 primary locations, 9 intermediate nodes, and 4 final destinations. Formulated as a min-cost flow problem with capacity constraints, minimizing travel time while maximizing throughput. Applied to non-combatant evacuation operation (NEO) planning scenarios.
Replicated and extended three published cryptanalysis techniques using neural networks and genetic algorithms for substitution cipher decryption. Achieved 92% decryption accuracy at 3.5× the speed of traditional frequency analysis. Combined neural sequence modeling with adaptive-mutation evolutionary search.