WebGPU-Spy is a hardware-based michroarchitectural attack that takes advantage of the low-level nature of the WebGPU API. It uses a cache occupancy channel to determine which website the victim is using.

ParticleGrid is a project to enable Deep Learning on Molecule datasets by leveraging the power of convolutional networks for the purpose of drug/material discovery. Convolutional neural networks are behind nearly all of the biggest advancements in image processing, and have potential to do the same for molecules.

The idea is to take a SMILES string, which is a textual representation of a molecule, and turn it into a 3D grid of atom densities. Using existing state-of-the-art convolutional techniques on such a grid is then very straightforwad. In order to facilitate maximum efficiency for property prediction, I worked to enhance the performance of this process considerably. I was able to achieve a best-case speedup of 100,000x over my colleague's initial numpy implementation.

The overall process is is to go from SMILES string (a) to 3D grid (f). For more info, take a look at the paper! We were able to generate these grids, and then create a model using Pytorch to predict the homo-lumo gap of a material.

Everyone knows dark mode is best. Or something. But what if... you had the option to choose? Well, some brilliant people have decided to make that dream a reality, and I decided to challenge myself to do it using only CSS, without using any JS.

Mobile pop-out menus are super nice, but you'd expect to need JS for something that complex. But what if... you didn't? Similarly to my no-JS dark mode, I made a mobile menu that doesn't need any JS to function, and it even has a fancy animation! It doesn't even use calc()!

Svelte is awesome, and it's how I created this website. This is a writeup on how I got Svelte/SvelteKit working on GitHub pages. Turns out it's very easy to do! I wanted a quick guide that shows all the steps without too much fluff, so that's what this is.

I wanted to understand how backpropagation works, so I built a backpropagation algorithm in python using numpy. It's not exactly high-speed, but it gets reasonable results for a basic feed-forward network on the MNIST dataset.