Leonardo Fabyan Ortega Rivera
_{Department of Computer Science · Universidad de Guadalajara (CUCEI)
Guadalajara, Jalisco, México}

Abstract — I am a Computer Engineering graduate candidate at Universidad de Guadalajara (100% coursework completed, GPA: 91.42/100) specializing in deep learning mathematical foundations, spatiotemporal modeling, and robust computational systems. My work is anchored in white-box engineering: understanding and deriving neural architectures from mathematical first principles—such as analytical Jacobians via tensor contractions (torch.einsum) without automated differentiation (0.00% MAPE vs. native PyTorch)—auditing and reproducing peer-reviewed scientific literature (uncovering published equation errata in Elsevier 2024), and enforcing rigorous software craftsmanship (strict static typing with Mypy, deterministic environments with Astral uv, and automated security CI/CD). Actively seeking early-career opportunities as a Junior Machine Learning Engineer, Computer Vision Engineer, or Systems / Python Developer.

White-Box Deep Learning and Analytical Gradients from First Principles

• First-Principles Derivation: Derived and implemented backpropagation, multi-dimensional tensor transformations, and exact analytical Jacobian derivatives without relying on torch.autograd.backward().
• Einstein Summation: Leveraged torch.einsum for tensor contractions and vector-Jacobian products (VJPs), achieving 0.00% Mean Absolute Percentage Error (MAPE) verified numerically against native PyTorch functional layers: $$\text{MAPE} = \frac{1}{n} \sum_{i=1}^n \left| \frac{y_{\text{analytic}} - y_{\text{autograd}}}{y_{\text{autograd}}} \right| = 0.00%$$
• Pedagogical Rigor: Authored companion mathematical derivations in MyST Markdown documenting computational graph topologies, gradient flow, and memory layouts.

Scientific Reproduction & Mathematical Audit — Elsevier (Cognitive Systems Research)

• Formal Replication: Independently reproduced the evolutionary multi-agent model of Sandoval-Arrayga et al. (Cognitive Systems Research, Elsevier 2024), confirming that multiscale temporal processing ($a=3$) yields up to a +151.8% fitness advantage over synchronous baselines across 4,500 generations.
• Mathematical Audit & Errata: Uncovered critical arithmetic inconsistencies in the published publication equations (e.g., $0.0833%$ vs. $0.833%$), documented formally in typos.md.
• Empirical Rigor: Designed and executed a 12-run parallel ablation suite (~10.6 hours compute) verifying fitness distributions, behavioral convergence, and evolutionary stability.

Physics-Guided Machine Learning & Dark Matter Substructure Detection (ML4Sci Challenge)

• Physics-Informed Architecture: Developed a Physics-Informed Neural Network (PINN) leveraging Radial Symmetry Residuals (RSR), a parameter-free mapping that isolates non-axisymmetric perturbations from the idealized Singular Isothermal Sphere (SIS) gravitational lensing model.
• Modified Vision Backbone: Adapted ResNet-18 to accept multi-channel tensors (raw lensing flux concatenated with RSR maps); conducted Bayesian hyperparameter optimization with Optuna over 100 epochs, achieving 93.76% validation accuracy and a macro-average 0.9915 One-vs-Rest ROC-AUC.
• Interpretability & Manifold Analysis: Validated physical feature localization via Grad-CAM saliency maps (localizing on genuine mass subhalos rather than global flux artifacts); performed topological manifold analysis (Silhouette: $-0.0022$, Davies-Bouldin: $109.92$) demonstrating severe linear non-separability, formally justifying deep convolutional representations over shallow methods.

Spatiotemporal Modeling & Multi-Frame Video Prediction

• Architecture Benchmarking: Engineered and evaluated spatiotemporal architectures (RCNN2d, Conv2dGRU, and non-recurrent baseline CNN) for sequence-to-sequence video prediction on Moving MNIST using 5D tensors $(T, B, C, H, W)$.
• Multi-Objective Perceptual Loss: Formulated a composite loss function balancing pixel-level error (MSE) and structural perceptual coherence (SSIM) with autoregressive scheduled sampling: $$\mathcal{L}{\text{total}} = \mathcal{L}{\text{MSE}} + \lambda , (1 - \text{SSIM})$$
• Recurrence Analysis: Analyzed memory retention vs. gradient attenuation in recurrent convolutions across variable rollout horizons.

Production-Grade Computer Vision & Automated CI/CD Lifecycle

• Vision Backbone: Transfer learning with MobileNetV3 for 6-class solid waste sorting, integrating deterministic CUDA seeding and multi-class One-vs-Rest ROC-AUC evaluation.
• Modern Python Tooling: Managed strictly through Astral uv for reproducible virtual environments and lightning-fast dependency resolution.
• Engineering Craftsmanship: Automated CI/CD pipelines in GitHub Actions enforcing strict static typing with mypy, formatting/linting with ruff, and automated security vulnerability auditing with bandit.

Algorithmic Problem Solving & Competitive Programming

• Algorithmic Base: Rigorous problem solving in C++17 and Python focusing on core data structures, greedy approaches, dynamic programming, binary search, and graph traversals.
• Structured Explanations: Every problem is documented with step-by-step reasoning, asymptotic complexity analysis ($\mathcal{O}(N \log N)$ time, $\mathcal{O}(1)$ space), and edge-case validation.

• 🎓 B.S. in Computer Engineering — Universidad de Guadalajara (CUCEI)
  • Coursework: 100% completed · GPA: 91.42 / 100 (Academic standing in high percentiles).
  • Specialization: Machine Learning, Numerical Methods, and Systems Programming.
• 🛰️ Scientific Research Exchange — Programa DELFÍN (Colombia)
  • Institución Universitaria de Envigado (2025): Architected the relational schema (OUL-DB) to catalog multi-spectral FITS astronomical images and satellite telemetry.
• 🌌 Latin Lunar Lobby Congress — Universidad Complutense de Madrid
  • Speaker on Dynamic NeRFs for Exospheric Reconstruction: 4D temporal parameterization of non-rigid diurnal particle density shifts for the OUL satellite mission.
• 💼 Industry Experience — Power Industrial (Software & Systems Developer)
  • Engineered warehouse inventory management software handling 3,600+ industrial SKUs with Python and SQLite3 (WAL mode) across multi-workstation local area networks.

If you find these implementations or mathematical notes useful in your work or research, please consider citing:

@misc{ortega2026whitebox,
  author       = {Ortega Rivera, Leonardo Fabyan},
  title        = {White-Box Deep Learning and Computational Systems: Mathematical Foundations, Empirical Falsifiability, and Reproducible Engineering},
  year         = {2026},
  publisher    = {GitHub},
  journal      = {GitHub Profile Repository},
  howpublished = {\url{https://github.com/PilotLeoYan}}
}

_{Open to technical discussions, collaborations, and engineering roles:}
leofabyano@gmail.com · LinkedIn · Curriculum Vitae (PDF)