Played a key role in developing a highly requested internal tool while working in the Windows Servicing and Delivery team (Experiences and Devices).

The IRIS-HEP Analysis Systems group is investigating whole-analysis differentiability to improve analysis optimization (grad-hep). However, not all operations in Awkward Array can be differentiated, so an analysis that uses this library can’t take advantage of this technique. There are several popular machine learning frameworks that make use of auto-differentiation. Out of these, Tensorflow, PyTorch and JAX are the most popular ones. My work involved computing derivatives for operations in Awkward Arrays, using these libraries so that all functions containing Awkward Arrays can be differentiated by them. The ​grad-hep group of IRIS-HEP is primarily focused on end-to-end analysis, and they use JAX as their primary library for auto-differentiation. Awkward Arrays and Uproot are becoming a standard within the particle physics community and without having derivatives of Awkward Array operations in place, the entire idea behind whole-analysis differentiability(​neos​) would be unable to proceed. During my fellowship I worked extensively to enable JAX to differentiate through functions containing Awkward Arrays.

Over the summer I worked on a project to support a GPU / CUDA-enabled backend for Awkward Arrays. With such a backend, the library now has the unique capability of processing complex, JSON-like data that is wholly resident on GPUs, in an efficient way, without specialized compilation or even leaving the Python prompt. Deployment was one of the major parts of this project and the we now have a seamless transition from the CPU backend to the GPU backend, with the dependent projects having to make minimal changes to accomodate for the new backend.