Working at Augment Code, and seed companies

TBA

Creating AI to assist warehouse order picking without manual labelingLed team that created the PopSignAI dataset, the world's largest isolated sign language dataset and first to use smartphones ("PopSign ASL v1.0: An Isolated American Sign Language Dataset Collected via Smartphones", NeurIPS 2023)Developed and deployed Android data collection app through DPAN Deaf members and data QA engine through reviewersDeveloping an isolated sign language recognition model using visual transformers to be used in mobile dictionariesCo-developer of the educational game CopyCat and co-author of "CopyCat: Using Sign Language Recognition to Help Deaf Children Acquire Language Skills" (CHI 2021 SRC, 1st Place in Student Research Competition)Created fingerspelling recognizer that received 1st place poster at Georgia Tech's 2021 Undergraduate Research Symposium.

As a TA for Artificial Intelligence, I created exam questions, held office hours for a class of over 900, and designed a new homework assignment teaching about Hidden Markov Models in response to the rise in plagiarism from ChatGPT.

As a teaching assistant for Computer Organization and Programming, I have managed course schedules and resources vital to the functioning of the course (such as grades, assignments, and auxiliary tools including autograders). In addition, I have instructed to students twice weekly and facilitated the consolidation of course material.

I redesigned how Meta’s user data flow tracking system associates asset usage with products or features in order for Meta to better track how user data flows in legacy code and to better understand and protect user data. Previously, assets were associated with nodes representing features or products by adding an annotation in the code of the asset which contained one short name. My changes removed the former restrictive one short name to one node restriction and implemented new validation to ensure that only one short name in a node was primary (selected over other short names if one is needed). The change required performing an internal migration of over 100 thousand short names and maintaining stability while gradually rolling out these changes. The end result lowered technical debt by refactoring over 25 files of code to eliminate logic made redundant by these schema changes. In addition, it simplified the node and short name management user interface, making it easier for product managers to change short names and reducing the barriers to tracking privacy.

I built an incident analysis system using Kusto, Typescript, and C# to automatically analyze Azure Batch (HPC) incidents. I achieved this by analyzing over 40 incidents to identify most valuable areas to merit incident analysis implementation. This required collaborating with software engineers using the tool and other stakeholders to scope, refine, test, and deploy the analysis system.

I designed a feature recognition system utilizing BigQuery and WEKA, a machine learning toolkit, in order to find correlations between bugs and configurations of systems, thus enabling developers to more easily determine potential causes and solutions for various bugs in fuel systems. Developing the system required analyzing over 400 fuel configuration data files.