I’ve been programming since I was 11 years old. These days, I mostly write Python code. In the past, I enjoyed writing C/C++/CUDA (and dabbled in JavaScript).
Fairer Features
I think fairness in ML is of utmost importance. Fairer Features is a production modeling pipeline that, given an image dataset, can produce a detailed PDF report conveying the demographic make-up of the dataset (race, age, sex), as well as how each demographic group is depicted. The modeling pipeline consists of CNNs and Large-Language Models, employing novel techniques for obtaining demographic depiction information.
Originally, I planned on launching this code into a startup–I interviewed with YCombinator for their 2023W batch. However, I think I’ll end up open-sourcing it soon, in addition to releasing a paper.
DSRPytorch
I really liked the deep symbolic regression paper. The authors only released a TensorFlow implemenation–I wanted to use this method for my own research, and I prefer PyTorch. So, I read the paper and implemented it in PyTorch. There were some cool parts involved, such as batching of variable-length sequences and a sequence-to-PyTorch transpiler. The code is available here.
GPU Optimization Code
I wrote a conjugate gradient solver using CUDA. I employed some fun computational tricks, which enabled my code to outperform a naive GPU implementation by an order of magnitude: (1) Loop unrolling; (2) Various data optimizations involving both concurrency and access patterns; (3) Floating point precision starts out very coarse and gradually becomes more precise as the algorithm iteratively approaches its tolerance
D2 BioSoftware
Using OpenCV and Electron, I wrote an application for ingesting images of petri dishes and computing how many bacterial colonies are present. Under the hood, I use a variety of image-processing techniques like bilateral filtering, adaptive threshold, and various other transformations. Then, I binarize the transformed image and feed it into some blob detection code that I wrote.
This Software is currently used by the McGowan Institue of Regenerative Medicine at the University of Pittsburgh.