Research
I’m Raymond Dueñas, a fourth-year Ph.D. student in Computer Science and Engineering at UC San Diego, building systems for efficient AI and LLM inference on edge and heterogeneous CPU/GPU hardware. My current work includes FlowFi (fault resilience of ML accelerators), UnifiedSplitting (CPU–GPU co-inference on edge AI platforms), and Junkyard GPU, a low-cost cluster of repurposed consumer devices that pool their compute to run large language models locally. My undergraduate research follows below.
FlowFi
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
Advisers
Dr. Pat Pannuto
Associate Professor
Computer Science & Engineering
University of California San Diego
Computer Science & Engineering
University of California San Diego
Dr. Ryan Kastner
Professor
Computer Science & Engineering
University of California San Diego
Computer Science & Engineering
University of California San Diego
Abstract
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. An early version was 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.
Figures/Images
Project Paper
Autonomous Systems
Adviser
Dr. Adham Atyabi
Associate Professor
Computer Science
University of Colorado Colorado Springs
Computer Science
University of Colorado Colorado Springs
Abstract
Autonomous uncrewed aerial vehicles require the ability to navigate various environments without collision failures. These systems already serve essential roles in various fields, from entertainment to military applications. There is a desire to replace costly multi-sensor-based designs with a system based solely on computer vision. However, these systems suffer from varying success rates in object recognition and obstacle avoidance. In some cases, object recognition is too slow to avoid collision failure. Current state-of-the-art solutions for computer vision-based systems implement artificial neural networks or an algorithm written specifically for a particular task. This work proposes a paired convolutional neural network architecture for the execution of object recognition and obstacle avoidance in uncrewed aerial vehicles. This work anticipates the paired convolutional neural network architecture to produce a vision-based autonomous system with high success rates in obstacle avoidance and a low rate of collision failures while maintaining competitive flight speed.
Figures/Images
Project Paper
Computer Vision for Analysis and Visualization of Climate Change Trajectory
Adviser
Dr. Kyu Han Koh
Associate Professor
Computer Science
California State University Stanislaus
Computer Science
California State University Stanislaus
Abstract
Modeling climate change has been an area of research since the 1970s. Since the 1970s, Climate Science has gained significant contributions as the field moves towards a better understanding of climate change. In 2022 alone, the NSF budgeted over $1 billion for research relevant to climate science. The development of climate data is essential for monitoring the climate system as well as detecting contributors to climate change, measuring the impacts of climate variability, and supporting an improved understanding of the climate system. This work aims to contribute to this body of research through the development of a novel methodology for creating accurate climate data sets based on image data. Utilizing computer vision on image data, this work intends to produce a climate data set with comparable values to data sets produced by NOAA and NASA. We expect the results of analysis on the developed data sets to provide additional evidence of climate change and global warming.
Figures/Images
Project Paper
Fitting a change point model for cardiovascular mortality in LA County
Adviser
Dr. Yanhong Wu
Associate Professor
Mathematics
California State University Stanislaus
Mathematics
California State University Stanislaus
Abstract
In this project, we analyze the weekly mortality of LA County from 1970 to 1979 due to cardiovascular complications. Our contributions are in two aspects. First, we find a better model fitting by using the change-point model instead of the linear trend model by treating the temperature and pollution as covariates, and the improvement is quite significant. Second, the time series analysis of the residuals shows that AR(2) gives a quite satisfactory fitting for the errors, and it can give better predictions. All the computations are carried in R, and R-codes and outputs are given.
Figures/Images
Project Paper
