CAPSA: A Model-Agnostic Framework for Risk-Awareness

As part of my work at Themis AI, I spearheaded the development of CAPSA, an open-source toolkit to automatically diagnose risk in ML models, regardless of the architecture or data modality. CAPSA can be used through a one-line addition to any training workflow. By leveraging shared feature backbones and composing risk metrics, we are able to achieve state of the art risk estimation in a more efficient and robust manner. Our paper, on which I am a co-first author, was published at the Robot Learning Workshop at NeurIPS ‘22.
talk | paper | code

6.4212 Final Project: TetrisBot

6.4212 is a graduate class that focuses on creating robotic systems that can automatically manipulate physical objects in unstructured environments. For our final project, we developed TetrisBot: an end-to-end robotic system that plays Tetris. We developed the perception, control, and gameplay systems that Tetris requires using Drake. This project won the Best Project Award for Fall 2022.
video | paper

6.8610 Final Project: Predicting Metadata from Song Lyrics Using NLP

6.8610 is MIT’s graduate NLP class. For our final project, we developed a method for automatically predicting song metadata (attributes such as danceability, energy, etc). Our algorithm shows high promise and is a step towards replacing Spotify’s crowdsourced metadata system, which can be highly inaccurate and incomplete.

6.869 Final Project: StyleAI: A Deep-Learning Powered Outfit Generator

6.869 is MIT’s graduate CV class. For our final project, we explored whether a model could ‘learn’ a sense of fashion. We did this by developing a novel contrastive learning approach to outfit generation, and found that outfits generated by the model were both plausible and creative.