Jose G. Perez, Ph.D.

Selected work

Satellite imagery, fluid forecasting, and atlas alignment.

Remote sensing · computer vision

Mapping debris-covered glaciers from satellite imagery

Debris-covered ice can look like the surrounding rock. I added terrain features and a velocity-based loss to improve segmentation of this difficult class.

Debris-covered ice IoU

Prior SOTA Boundary-Aware U-Net 35.94%

Our best model Complete physics-informed model 46.07%

+10.13 points · +28.2% relative

Clean ice IoU

Prior SOTA Boundary-Aware U-Net 68.17%

Our best model Full Physics terrain features 71.22%

+3.05 points · +4.5% relative

Ph.D. dissertation · 2025Physics-Guided Strategies for Enhancing Neural Networks Trained With Limited Data ↗View source ↗

Satellite image of glacier terrain — **Debris-covered ice**The surface resembles surrounding rock, but the prediction retains most of the glacier identified in the expert labels.
Orange: debris-covered glacierBlue: clean glacier ice

Model glacier prediction over satellite imagery — **Debris-covered ice**The surface resembles surrounding rock, but the prediction retains most of the glacier identified in the expert labels.
Orange: debris-covered glacierBlue: clean glacier ice

Precisionselected example90.4%

Recallselected example80.8%

Intersection over unionselected example74.4%

Scientific machine learning

Predicting fluid velocity beyond the training conditions

The models learned from a 250 m/s simulation and predicted flow at 300 m/s. I compared a bidirectional LSTM on its own with a physics-informed LSTM hybrid that combines the same LSTM with a Navier–Stokes-constrained network.

Horizontal velocity U

Baseline LSTM alone 1.7811

Our model Physics-informed LSTM hybrid 0.4677

73.7% lower MSE

Vertical velocity V

Baseline LSTM alone 8.6962

Our model Physics-informed LSTM hybrid 1.2794

85.3% lower MSE

ASME FEDSM 2022Physics-Informed Long-Short Term Memory Neural Network Performance on Holloman High-Speed Test Track Sled Study ↗View source ↗

How the model works

One forecast, informed by both data and fluid mechanics

Hover or focus a stage to follow its contribution

Reported mean-squared error

Prediction error ↓ lower is better

Horizontal velocity UVertical velocity V

No-change baselineUses the previous velocity field

U 5.455V 2.561

LSTM only

U 1.7811V 8.6962

Physics + LSTM

U 0.4677V 1.2794

Medical imaging · feature matching

Matching anatomy across rat-brain atlases

I built a SIFT and RANSAC method to compare internal tissue landmarks across the Paxinos & Watson and Swanson rat-brain atlases, then checked the matches against manual Bregma alignments.

Frontiers in Systems Neuroscience · 2018Computer Vision Evidence Supporting Craniometric Alignment of Rat Brain Atlases to Streamline Expert-Guided, First-Order Migration of Hypothalamic Spatial Datasets Related to Behavioral Control ↗View source ↗

Manual Bregma checks

Feature matches across four aligned pairs

Choose a pair to inspect how corresponding internal anatomy is linked across the two atlases.

Green lines: RANSAC-verified feature matches

RANSAC-verified feature matches between Paxinos and Watson level 68 and Swanson level 33 — **PW14 level 68 ↔ Swanson level 33** Same manual Bregma coordinate
**0 μm apart**

The pairs come from the paper’s manual Bregma alignment table. The correspondence views were regenerated with the project matcher; each green line is a RANSAC inlier.

Career

Selected experience.

Turned a multispectral classification workflow into a Prefect service for scheduled and on-demand inference
Migrated model workflows from TensorFlow to PyTorch and supported triplet-loss training for one-shot classification
Built Docker deployment workflows and GitLab CI pipelines for tests and artifact packaging
Built data-exploration and mission-dependency interfaces, including OpenAI-assisted workflows

Improved debris-covered ice segmentation by 28.2% relative to the prior benchmark, reaching a 46.07% overlap score (IoU) after adding terrain features and velocity-aware loss
Improved U-velocity by 73.7% and V-velocity by 85.3% over an LSTM-only baseline by incorporating Navier–Stokes constraints
Designed controlled ablation and reproducibility studies and maintained Ubuntu/NVIDIA GPU servers
Built OpenCV, SIFT, and RANSAC rat-brain atlas alignment software, published in Frontiers

Supported Data Structures, Computer Vision, Deep Learning, and Machine Learning at undergraduate and graduate levels
Delivered ML instruction for US Army learners at White Sands Missile Range

Capabilities

Models, experiments, and systems.

Computer vision

Segmentation, registration, and one-shot classification for scientific imagery.

Scientific ML

Physical features, constraints, and controlled experiments for limited data.

ML systems

GPU workflows, orchestration, containers, CI/CD, and model interfaces.

Working toolkit

Tools and technologies

Core

Python
PyTorch
OpenCV
Linux

Shipped

Docker
Prefect
PHP
Vue.js
MySQL
TypeScript
Rust

Applied

C++
C#
Java
GitLab CI/CD
CUDA / GPUs

Education & recognition

Training and recognition.

Education December 2025

Ph.D. in Computer Science

University of Texas at El Paso

Thesis: Physics-Guided Strategies for Enhancing Neural Networks Trained With Limited Data. Advisor: Dr. Olac Fuentes.

Award 2023

Google-CAHSI Dissertation Award

Google / UTEP

$25,000 dissertation award.

Education May 2018

B.Sc. in Computer Science

University of Texas at El Paso

Cum Laude. Minors in Biomedical Engineering and Mathematics.

Award 2015-2018

BUILDing Scholars Traineeship

National Institutes of Health

NIH-funded full-ride undergraduate research traineeship.

Research record

Peer-reviewed publications.

In progress · 2026 Physics-Guided Glacier Segmentation with Static Terrain Features and Dynamic Velocity Constraints

Contact

Building with difficult visual or scientific data?

I’m looking for US-remote roles in applied ML, computer vision, scientific ML, and applied science. Tell me what the model needs to do and what makes the data hard.

Email me about a role Download résumé LinkedIn

I build machine learning systems for real-world data.