Projects/Streams

Here is my GitHub and a YouTube playlist describing some of my projects.
  1. Guaranteeing JSON output from GPT-2 (video, repo)

    2. I force GPT-2 to generate output in JSON, and do this by fixing tokens to be parts of the JSON schema (the "keys") and then only generating the "values" of the key:value pairs. Compared with other methods which make repeated calls to the LLM and dynamically build the prompt, we only make *one* call to the LLM seeing as we have direct access to the model.

  2. Sherlock Holmes Eval (repo, video)

    3. Inspired by an episode of the Dwarkesh Podcast, this is an eval that tests LLMs ability to determine the culprit in Sherlock Holmes murder mysteries.

  3. Training algorithm for regularized models on arbitrarily large data sets (video/arXiv/repo)

    Eliminates requirement of storing O(n) quantities in memory, and allows for training data to be stored in distinct sites for privacy concerns.

  4. Debiased high-dimensional logistic regression (video/arXiv/repo)

    Conjecture to debias model in setting where both the number of features and observations are asymptotically increasing.

  5. AI-powered flashcard generator and spaced repetition study tool (repo)

    6. This is a web app which uses a Django backend, React frontend, and the OpenAI API that allows users to upload documents and be turned into flashcards which can be studied using spaced repetition memory systems.

  6. Europe Citadel Datathon 2021 (repo)

    7. This was a team event hosted by Citadel. The task was to find new insights about how Covid-19 spreads. We placed second, here is the final report