Raymond Dueñas

PhD Student | Computer Science and Engineering at UCSD

Open to Summer 2027 research & industry internships · r1duenas@ucsd.edu · LinkedIn

Raymond Dueñas in cap and gown with McNair Scholar and LSAMP stoles at his CSU Stanislaus graduation

Hello!

I’m Raymond Dueñas, a fourth-year Ph.D. student in the Department of Computer Science and Engineering at the University of California San Diego, where I’m fortunate to be co-advised by Dr. Pat Pannuto and Dr. Ryan Kastner.

My research focuses on making AI hardware both reliable and efficient. As machine learning moves into accelerators at the edge, from scientific instruments to embedded systems, a single hardware fault can silently change a model’s answer; I study how accelerator design choices, such as how data flows through a systolic array, shape resilience to those faults. My additional focus is on making AI inference fast on the hardware people already have: edge devices, heterogeneous systems that pair CPUs with GPUs, and other available processors. I study how deep learning and large language model (LLM) workloads behave across processors and devices. I build systems that place workloads where they perform best, and tools that inform those placement decisions so the workloads run with optimized resilience.

Prior to joining UCSD, I completed my undergraduate studies at California State University, Stanislaus, with dual majors in Mathematics and Computer Science, graduating as an Honors student. During my undergraduate career, I was honored to be a McNair Scholar, a Cal-Bridge Scholar, and an NSF Louis Stokes Alliances for Minority Participation (LSAMP) Scholar. I was also nominated as a 2022–2023 CSU-LSAMP PROUD Scholar for Outstanding Academic Performance and Outstanding Service. I was especially honored to receive the Outstanding Scholar in Mathematics award for my graduating class, recognizing my achievements in the Bachelor of Science in Mathematics program.

As an undergraduate, I worked with Dr. Kyu Han Koh in the Innovative Learning and Design Lab on computer vision applications for climate data analysis and visualization. I also collaborated with Dr. Adham Atyabi at the University of Colorado Colorado Springs on research utilizing computer vision and deep learning for autonomous drone flight with optical flow for obstacle avoidance.

At a glance

Research areas: ML systems · ML accelerator reliability and fault injection · efficient LLM and deep learning inference · edge and embedded AI · heterogeneous CPU/GPU computing · distributed inference · GPU performance profiling
Tools: Python, C++, CUDA, PyTorch, Hugging Face Transformers, TensorRT, llama.cpp, Vulkan, Docker, Kubernetes (K3s), Linux, Git, NVIDIA Nsight Compute/Systems
Infrastructure: cluster system administrator for the lab’s ~40-device ARM64 edge-AI research cluster: Kubernetes (K3s), networking, OS images, and access management for student researchers
Leadership: led research projects with a team of 20 students; mentored four undergraduate researchers
Awards: Best Poster Award, FastML 2026 (FlowFi)
Fellowships: Cal-Bridge Fellow (2023–present); UCSD CSE Department and Rotational Fellowships (2023)

Open to research and industry internships for Summer 2027. Email: r1duenas@ucsd.edu · LinkedIn

Lab

Dr. Pat Pannuto

PatLab

Dr. Ryan Kastner

Kastner Research Group

Hand holding a glowing light bulb surrounded by gear icons, a decorative image for research ideas

Projects at a Glance

FlowFi: Fault Resilience of ML Accelerators

As machine learning moves into hardware at the edge (detector triggers, satellites, embedded systems), a single faulty bit can silently change a model’s answer. FlowFi is a cycle-accurate fault-injection simulator for systolic-array ML accelerators that measures how the choice of dataflow shapes resilience to hardware faults, across integer and low-precision floating-point formats.

★ Best Poster Award, FastML 2026. Advised by Dr. Ryan Kastner. Manuscript in preparation.

UnifiedSplitting: CPU–GPU Co-Inference on Edge AI Platforms

Edge AI platforms such as NVIDIA Jetson place the CPU and GPU on a single shared memory. UnifiedSplitting explores how to divide deep learning inference across those processors to get more throughput from the same device, with no changes to the model. I build the inference pipelines in Python and PyTorch and profile them with NVIDIA Nsight Systems and Nsight Compute. Presented as a lightning talk and poster at the UCSD CNS Research Review (May 2025). This research is co-advised by Dr. Pat Pannuto and Dr. Ryan Kastner in the Department of Computer Science and Engineering at the University of California San Diego.

Line chart of ResNet-18 inference throughput by batch size, comparing full-GPU and UnifiedSplitting runs

Diagram contrasting a discrete GPU with separate CPU and GPU memory and an integrated GPU sharing system memory

Learn More

Junkyard GPU: Distributed LLM Inference on Repurposed Devices

Building a low-cost cluster of repurposed consumer devices that pool their compute to run large language models locally, without a data center.

Optical Flow for Obstacle Avoidance in Autonomous Drone Flight

Utilizing a monocular quadcopter, this project aimed to implement a pair of convolutional neural networks (CNNs) to produce an autonomous flight controller that successfully navigates the drone to its destination. The first CNN produces optical flow values. Accepting these values as inputs, the second CNN makes the optimal directional flight decision to get to its destination safely. This work was started under the advisement of Dr. Adham Atyabi during the summer of 2022 at the University of Colorado Colorado Springs.

Diagram of the drone pipeline: optical-flow CNN and path planner feed a motor-controller CNN that outputs direction commands

Grayscale office view overlaid with green optical-flow vectors and a 3x3 grid for drone obstacle avoidance

Small Tello quadcopter drone with four propellers, used in the optical-flow obstacle avoidance project

3D plot of blue and red data points above a gray reference plane, from the drone obstacle avoidance work

Learn More

Computer Vision for Analysis and Visualization of Climate Change Trajectory

Implementing computer vision practices and the OpenCV library on images rendered from data sets containing fifty years of climate information, this project aimed to track the trajectory of the observable effects of climate change. This work was conducted under the advisement of Dr. Kyu Han Koh as a project of his Innovative Learning and Design Lab.

Color-coded temperature map of the western United States, hottest in California's Central Valley and southern deserts

Climate CV steps: western U.S. temperature map, a binary mask of the hottest areas, and those areas isolated in orange

Learn More

Fitting a Change-Point Model for Cardiovascular Mortality in LA County

Implementing time series analysis and change-point regression techniques on a decade of Los Angeles County mortality data, this project aimed to identify structural breaks in cardiovascular death patterns while accounting for environmental covariates. The methodology employs R statistical computing to fit autoregressive models with temperature and pollution variables, demonstrating that change-point models provide significantly superior fitting compared to linear trend approaches. Through systematic F-value optimization across 507 potential change points, the analysis reveals a distinct mortality shift at week 262, with AR(2) error modeling achieving robust residual correlation below 0.10 tolerance thresholds. This work was conducted under the advisement of Dr. Yanhong Wu in the Department of Mathematics at California State University Stanislaus.

Scatterplot matrix of LA County cardiovascular mortality, temperature, and particulate levels with red trend annotations

Autocorrelation (ACF) plot of the mortality model's residuals by lag, annotated with a red decay curve and highlighted point

Learn More