Samantha Andow

Jun 2021-Present

Still at Meta, transfered to PyTorch team (and to New York)
I work mostly on functorch ()most functorch PRs, including designing and implementing our randomness API)
Other than that, I've also developed the ExpandedWeights API which let users compute per sample gradients. This was used by the Opacus team to make their library twice as fast
Finally, I built and improved the TorchDispatchMode system. This lets users add behavior on all calls, which is necessary for tracing by torchdynamo and made debugging easier for developers

At Meta, developed a differentiable programming framework in Kotlin
I worked on compiler plugin that adds a function in the IR that computes the backwards formula without needing a tape-based implementation
After we decided that we shouldn't move forward with this plugin, I switched to optimizing our runtime, tape-based system

For a clinic project, I worked with Steelcase to develop a mobile system to detect wood veneer color
We built a box with an onboard Raspberry Pi and camera to eliminate outside light and control the environment when doing the color detection
In doing so, I also rewrote their original color detection system to be significantly faster
Writeup from Harvey Mudd College

As a SWE intern at Microsoft, I built a prototype that showed Kubernetes could run in a distributed system
It ran each part of the Kubernetes system in a different Docker container on a different machine. Then, when one of the instances went down, it would be able to self replicate on the same Docker container

As a research assistant at Harvey Mudd College, I developed a new middle school CS curriculum called MyCS
Rewrote much of the previous curriculum and taught it to 3 different sets of teachers who use it in their classrooms