Jose G. Perez, Ph.D.
Ph.D. ML Engineer · Computer Vision & Scientific ML
I help teams turn uncertain research problems into working software: PyTorch models, reproducible experiment pipelines, Docker/Prefect services, and interfaces people can actually use. My edge is the combination of Ph.D.-level scientific ML, production-oriented APL experience, and long-running full-stack open-source engineering.
Selected work
Built on an existing glacier-mapping U-Net baseline, then added DEM-derived physics features, ITS_LIVE velocity constraints, dataset-fusion pipelines, and 85+ controlled experiments for Landsat 7 imagery of the Hindu Kush Himalayas.
Key result
46.07% debris-covered ice IoU (28.2% relative improvement over the dissertation benchmark). Combined physics-informed data augmentation with velocity-aware loss.
Top contributor to a long-running multiplayer strategy game server, spanning PHP backend logic, Vue UI, database migrations, Docker tooling, CI, replay infrastructure, and experimental Rust services.
Key result
Shipped player-facing and infrastructure work across replay streaming, S3/MinIO storage, hotseat mode, keyboard UX, player-box redesign, PHP 7.4 migration work, and production bug fixes.
Hybrid two-branch network combining LSTM with Physics-Informed Neural Network for predicting fluid flow fields from limited simulation data (2D incompressible Navier-Stokes on Holloman High-Speed Test Track CFD).
Key result
73.7% improvement in U-velocity and 85.3% improvement in V-velocity over LSTM-only approaches. Published at ASME FEDSM 2022.
Cross-atlas registration for mapping rat brain histological data between Paxinos-Watson and Swanson reference spaces using SIFT feature extraction and RANSAC geometric alignment.
Key result
Reduced manual burden of migrating spatial data between incompatible atlas editions. Published in Frontiers in System Neuroscience.
Core expertise
Landsat 7 satellite image segmentation and classification, one-shot learning with triplet loss, SIFT/RANSAC registration, U-Net with boundary-aware learning.
PINNs with Navier-Stokes, hybrid LSTM+PINN architectures, physics-informed data augmentation from DEM features, velocity-aware loss functions.
Prefect orchestration, Docker/Docker Compose, GitLab CI/CD, GPU server administration, multi-source geospatial data fusion pipelines.
Experience
May 2023 -- Dec 2024
Johns Hopkins University Applied Physics Laboratory (APL)
May 2018 -- Dec 2025
University of Texas at El Paso (UTEP)
Fall 2018 -- Summer 2025
University of Texas at El Paso (UTEP)
June 2016 -- Aug 2016
University of Texas at Austin
Education & recognition
University of Texas at El Paso
Thesis: Physics-Guided Strategies for Enhancing Neural Networks Trained With Limited Data. Advisor: Dr. Olac Fuentes.
Computing Alliance of HSIs
Conference travel grants supporting research presentation and participation.
Google / UTEP
$25,000 dissertation award.
University of Texas at El Paso
Cum Laude. Minors in Biomedical Engineering and Mathematics.
National Institutes of Health
Full-ride undergraduate biomedical research scholarship.
Publications
Preprint (in preparation) / 2026
ASME Fluids Engineering Division Summer Meeting / 2022
ASME Fluids Engineering Division Summer Meeting / 2022
SPIE Defense + Commercial Sensing Symposium / 2021
Frontiers in System Neuroscience / 2018
Research-to-production pipeline
Step 1
Literature review, problem formulation, physics-guided experimentation, publication.
Step 2
PyTorch, U-Net, LSTM, PINNs — training, hyperparameter sweeps, ablation studies across 85+ configs.
Step 3
Docker, Prefect orchestration, GitLab CI/CD, GPU server administration, experiment tracking.
Step 4
Research-to-production handoff, automation pipelines, reproducible experiment workflows, documentation.
What I build
Segmentation, classification, registration, and multispectral image analysis for teams that need a credible first model and evidence-backed next steps.
Physics-informed and limited-data modeling for PDE-constrained prediction, remote sensing, geospatial imagery, and research-heavy product bets.
Turn notebooks and lab pipelines into Dockerized, scheduled, tested workflows with Prefect, CI/CD, GPU environments, and usable interfaces.
Skills
Languages
ML & CV
MLOps & Infra
Domains
Writing & documents
Contact
Looking for help with computer vision, scientific ML, or research-to-production ML systems? Ph.D. in computer vision and scientific ML with production experience at JHU Applied Physics Lab.