Ryan Burn

I wrote a high performance tracing library in C++. The library makes use of multiplexing IO, arena allocation, and lock-free data structures to achieve high throughput tracing at a low CPU cost. It's used by Envoy, NGINX, and financial service customers that run high concurrency, low latency applications.

I developed a Python native module (written in C++) for building warped linear regression models. Warped linear regression extends ordinary least squares by adding an additional step where target values are transformed via a monotonic function to maximize likelihood. The additional step makes it more general purpose than OLS and capable of fitting models with non-normal error distributions.

I developed software that optimizes regularization parameters for ridge regression for certain cross-validation metrics. By efficiently expressing regularization as an optimization problem where both values and gradients are computed. The software is able to handle multiple regularizers and can often provide better results than what can be achieved with a traditional grid search.

I created an image stitching application based off of an IEEE research paper to correct positioning errors in overlapping images taken from a microscope. The application scored how well overlapping regions aligned with the zero-mean normalized cross correlation metric and applied a fast fourier transform to efficiently compute the metric for a range of translations. It then optimized the scoring metric across image placements using a dynamic programming algorithm. The impact of the program was to eliminate a time-consuming process in which images were manually adjusted by hand.

I worked as a lead developer to build an auction system for the Bloomberg Terminal used by central banks and primary dealers to buy and sell government securities. The application featured a real-time monitor for bids and customizable bid allocation strategies. I worked both on the back end that used C++ and SQL and the JavaScript front end.

I extended the GCC C++ compiler to better support the Cilk parallelization language extensions. I modified the compiler's intermediate code generation to allow the cilk_spawn keyword to work with expressions involving constructor calls or user-defined type conversion operators.

For many regularized models, including logistic regression, the leave-one-out cross-validation can be efficiently estimated with an approximate leave-one-out cross-validation formula (ALO). I derived equations for computing the gradient and hessian of ALO and developed a process for optimizing ALO to tune hyperparameters.