Jan 2024 – Jan 2025 · 1 yr · Remote

Aug 2022 – Jan 2024 · 1 yr 5 mos · Remote

Jun 2020 – Aug 2020 · 2 mos · Boston, Massachusetts, United States

• Collaborated on an internal Python package called PyAnzo which serves as a programmatic interface between users and Anzo, CSI’s graph-driven, modern data discovery and integration platform
• Road-mapped 1.0 and 1.1 package releases in weekly developer sessions
• Discovered bugs with the existing code-base and implemented bugfixes & performance improvements
• Devised new features that would help PyAnzo fit effortlessly into existing client workflows
• Prioritized documentation and testing infrastructure to accelerate a stable 1.0 release to CSI customers
• Directed multiple department and company-wide presentations to inform and train other employees on new features and technologies

Sep 2019 – Jul 2022 · 2 yrs 10 mos · Cambridge, Massachusetts, United States

• 𝗕𝗲𝘀𝘁 𝗽𝗮𝗽𝗲𝗿 𝗮𝘄𝗮𝗿𝗱: NeurIPS Cooperative AI Workshop (2021)
• Developing a new approach to Reinforcement Learning called Theory-Based Reinforcement Learning which performs Bayesian inference to learn probabilistic generative models expressed as programs for a game-engine simulator
• Increasing explainability of the model to enable sharing of key insights with human players and other models
• Coordinating large-scale web experiments (n > 200) to understand how humans can play through a game and then decide on the most relevant information to share when teaching the game to a new player and working to imbue this ability in our Theory-Based Reinforcement Learning model
• Implementing message passing in a symbolic manner that allows performant models to pass specific pieces of knowledge to newly-initialized model agents in order to dramatically improve learning outcomes
• Formulating methods and metrics for the model that allow it to select which bits of information are the most relevant/important to pass along to the new novice agents
• Applying state of the art NLP techniques (including GPT-2 fine-tuning and GPT-3 prompt engineering) to classify, understand, and parse open-ended participant messages and relate message content to game performance

Sep 2018 – Sep 2020 · 2 yrs · Cambridge, MA

• Tested and refined computational models of knowledge transmission and cultural learning
• Developed large-scale, web-based human experiments (n > 100) on topics such as Function Learning (how do humans learn and teach function fitting with limited data) and Iterated Antonym Elicitation (how do humans generate antonyms when presented with a stimulus word)
• Iterated Antonym Elicitation: Constructed well-documented analysis pipelines that involved probing word embeddings (Glove, FastText, BERT) to better understand the encoded dimensions, devising methods for feature extraction and dimensionality reduction, and implementing logistic regressions and clustering to build a classifier of semantic relations (i.e. distinguishing synonyms from antonyms)
• Analyzed real-world data and devised methods of effective feature extraction, logistic regressions, dimensionality reduction and clustering in order to distinguish different semantic relations

Jun 2018 – Aug 2018 · 2 mos · Bangalore, India

• Collaborated with the Cedar Group from UC Berkeley on The Synthesis Project, an open-source database of inorganic synthesis recipes from scraped HTML journal articles
• Developed a robust HTML parser to create a database of more than 200,000 parsed Nature journal papers
• Coded a workflow that stitches together machine learning and natural language processing APIs in order to extract chemical equations/synthesis conditions from HTML articles