Hello, I’m Gavin

• Designed and implemented a novel machine learning approach using PyTorch MLPs to reconstruct high-resolution neural shape models from Diffusion Tensor MRI scans.
• Enhanced anatomical fidelity by building models that improved structural detail capture in reconstructed neural shapes.
• Built an end-to-end preprocessing and modeling pipeline in PyTorch and NumPy, including segmentation and normalization.
• Collaborated with a Stanford Medicine postdoctoral researcher, delivering reproducible, production-ready code.

• Created a Python tutorial notebook introducing Spiking Neural Networks (SNNs).
• Demonstrated benefits of SNNs for reducing model storage and energy usage.
• Collaborated with researchers developing open-source tools for SNN model conversion and optimization in PyTorch.

Research conducted during a Stanford Medicine REU at the Radiological Sciences Laboratory, investigating whether neural implicit representations can accurately reconstruct muscle Diffusion Tensor Imaging data. The project developed a two-stage approach: first, a neural shape model (occupancy network) that learns to classify 3D coordinates as inside or outside a muscle segmentation, then a neural vector field that jointly predicts occupancy, principal diffusion direction, and diffusion scalars from spatial input. The occupancy network exceeded accuracy targets (DSC 0.929), while the vector field reconstruction revealed open challenges in capturing directional information — honest results that motivated ongoing work in the lab. Built end-to-end in PyTorch with SimpleITK for medical image I/O and custom SIREN-style architectures.

A clinical variant filtering and prioritization pipeline for Whole Exome Sequencing data from a NICU heterotaxy patient. The pipeline processes VEP-annotated variant files through a systematic filtering cascade — transcript selection (MANE Select or canonical), population frequency filtering against gnomAD, functional impact assessment, and ClinVar pathogenicity tiering — to narrow thousands of raw variants to a manageable set of clinically actionable candidates. Includes deep interrogation of DNAH5 (a key ciliary gene in heterotaxy) with zygosity stratification and SIFT/PolyPhen scoring, plus screening against a curated panel of 25 heterotaxy-associated genes spanning ciliary, signaling, and emerging candidate categories. Developed for a real clinical case as a prototype for potential commercialization.

A computational genomics research project investigating the role of the JAK2 V617F mutation as the primary driver of Polycythemia Vera, a rare myeloproliferative neoplasm. The pipeline integrates differential gene expression analysis, KEGG pathway enrichment, pre-ranked GSEA, and Spearman correlation matrices to build the case that JAK2 drives PV through constitutive protein-level activation rather than transcriptional upregulation — explaining why it doesn't appear among top differentially expressed genes. Analysis of MPN patient data from GEO reveals significant enrichment of immune/inflammatory and cell cycle pathways, with strong JAK2-STAT co-expression and active SOCS/CISH feedback. Includes exploratory WGCNA and course assignments covering Poisson/NB count modeling, doublet detection, and batch correction.

Six core bioinformatics and machine learning algorithms implemented from first principles in Python and NumPy. Covers the full spectrum from unsupervised learning (K-means on MNIST digits, PCA/MDS for dimensionality reduction, NMF for population structure analysis on dog genotype data) to statistical genomics (permutation testing for interval overlaps, differential expression with log2 fold change) to probabilistic sequence models (HMM with Viterbi decoding for detecting runs of homozygosity from VCF data). Each implementation avoids library abstractions to demonstrate understanding of the underlying mathematics — eigendecomposition, log-space computation, convergence detection, and backtracking algorithms.

A complete multi-layer neural network implemented from scratch in NumPy, with every component hand-coded: forward propagation through ReLU-activated hidden layers, backpropagation via chain rule gradient computation, softmax output with cross-entropy loss, mini-batch stochastic gradient descent, and learning rate scheduling. No deep learning frameworks used — the goal was to understand exactly what happens inside a neural network during training. Systematic experiments across architectures (2–4 hidden layers, 128–512 neurons) and hyperparameters (learning rates, batch sizes) culminated in 97.46% test accuracy on the full MNIST dataset.

A progressive portfolio of scientific visualizations built entirely in low-level Matplotlib — no seaborn, no plotly, no high-level wrappers. Each piece tackles a different bioinformatics visualization challenge: scatter plots with marginal histograms for gene expression, t-SNE cell-type clustering, beeswarm plots with custom collision detection, information-theoretic sequence logos, circadian expression heatmaps, and a full genome browser with gene model rendering (GTF), read alignment packing (PSL), and per-base coverage histograms. Demonstrates the ability to produce figures that meet journal publication standards with precise multi-panel layouts, custom color maps, and domain-specific plot types.

A suite of Python tools for core bioinformatics tasks, built progressively across a full course. Includes a reusable sequenceAnalysis module with FASTA parsing, codon counting, and GC content analysis; a ProteinParams class computing molecular weight, isoelectric point, and extinction coefficients; an ORF finder spanning all six reading frames with configurable start/stop codons; a tRNA unique subsequence identifier using set-based comparison; and a DNA storage degradation simulator modeling seven damage types with literature-derived error rates. Each tool emphasizes clean OOP design and reusable module architecture.

A transfer learning pipeline for satellite image classification under extreme data scarcity — just 10 labeled training samples per class. Built on a frozen pretrained ResNet-50 backbone with a custom classifier head incorporating Global Average Pooling, Channel Squeeze dimensionality reduction, and a Residual MLP block. Training pipeline includes aggressive data augmentation (random rotation, horizontal/vertical flips, color jitter, random erasing), cosine annealing with warm restarts, and label smoothing. Achieved 87.64% accuracy on EuroSAT and placed 3rd on the class Kaggle leaderboard out of approximately 50 students.

A two-part progression through core machine learning: Part 1 implements linear and logistic regression entirely from scratch in NumPy, with hand-coded MSE and binary cross-entropy loss functions, analytical gradient computation, batch gradient descent, IQR-based outlier removal, z-score normalization, and 5-fold cross-validation with ensemble inference on a wine quality dataset. Part 2 rebuilds and extends those concepts in PyTorch — feedforward neural networks as nn.Module subclasses, manual gradient descent without autograd, side-by-side visualization of Sigmoid/Tanh/ReLU activation functions, dropout regularization, and momentum-based SGD with velocity tracking. Together they show the full arc from implementing everything by hand to using modern frameworks.