Ashwin Devaraj

Multimodal LLMs at Meta Reality Labs

Working on paragraph-level simplification of text using Transformer-based language models, focusing on the medical domain Exploring the factuality of current simplification datasets and models

Used NLP methods to determine if tweets during natural disasters like Hurricane Harvey contain calls for help relevant to first responders. Trained non-neural classifiers (like SVM and AdaBoost) in Scikit-learn and deep convolutional neural networks in Pytorch with oversampling to achieve over 88% F1 scores on a highly unbalanced dataset.

Added new functionality to the vehicle data pipeline tool, such as filtering data points by specific features and displaying plots Wrote test cases for features I implemented and fixed minor bugs in the existing system

Cybersecurity and Technology Controls

I contributed build features to a GUI build tool (written in C# and Windows Forms) used by many teams across the firm. My contribution allows developers to configure and run multiple builds in parallel without interaction with the command line. I developed C# unit tests for simulation software and edited subtle build configurations to migrate old software to run on 64-bit Windows systems. As part of this migration, I ran static code analysis tools on the projects to resolve potential vulnerabilities in the code.