Personal Projects
Things I’ve built outside of my day job. For a complete list, please browse my GitHub.
Physics-Informed Neural Networks for Navier-Stokes Vortex Shedding
PINNs on the Schäfer-Turek cylinder benchmarks at Re=20 and Re=100, studying how training interventions interact and diagnosing a failure mode at higher Reynolds number.
[Code] [Paper]
Out-of-Distribution Detection for Plant Pathology
A systematic evaluation of OOD detection methods on real-world apple leaf disease images, spanning post-hoc softmax scoring, auxiliary objectives, energy-based fine-tuning, and constrained optimization.
[Code] [Paper]
GPT-2 from Scratch in PyTorch
A from-scratch implementation of GPT-2 in PyTorch Lightning, pretrained on FineWeb-Edu. Implements both the full 124M-parameter model architecture and a smaller 30M-parameter model. Includes evaluation scripts on HellaSwag.
[Code]
Analyzing Localization Abilities of Vision Transformers
Extends Grad-CAM beyond CNNs to vision transformers and to the object detection domain, the first work to do so to our knowledge. Compares where ResNet, DeiT, YOLOv3, and DETR attend when classifying and detecting objects.
[Paper]
CNN-Based Discovery of Transcription Factor Binding Sites in DNA
Deep learning models that identify protein-binding motifs in DNA, with saliency maps showing which bases drive each prediction.
[Code]
Unhinged: Settle It
An iOS app (currently on the App Store) for friend groups who settle arguments with data: create a prediction about anything and let everyone weigh in with their pick.
[App Store]Low-Rank Adaptation (LoRA)
A from-scratch implementation of LoRA fine-tuning, with notebooks walking through the math and an IMDB classification example.
[Code]PyTorch Experimental Framework
Reusable PyTorch templates for autoencoders, VAEs, GANs, transfer learning, and custom CNNs. Covers various optimizers, learning-rate schedulers, and hyperparameter sweeps tracked in TensorBoard.
[Code]