I take the view that varied experiences bring out the best of humanity: fostering creativity, bridging divides, and building intuition.
Growing up in the US midwest, my academic and work experience has taken me to California and back again, with a stopover for a posdoctoral appointment in Taipei.
An occasional physicist, data scientist, and developer, I always remain curious and try to weave together my accumulated knowledge. Most recently, I have been interested in exploring cognition from the point of view of algorithmic complexity, inspired by work surrounding program synthesis and the Abstraction and Reasoning Corpus.
Here are a few keywords describing where my career has ventured:
- Monte Carlo Methods
- Machine Learning
- Data Visualization
Where I’ve Learned and Worked
Senior Software Engineer @ Trace
March 2021 - October 2021
- First full-time employee
- Designed and built a configurable analytics backend in Python
- Built a React application to author and run analytics configurations
- Handled deployments to GCP, data ingestion, internal tooling
Neural Net Interpretability
I study neural network interpretability and fundamentals. Early work replicated results from OpenAI, Anthropic, Redwood Research, and others. Current efforts explore superposition via symmetric mixtures, network jamming, and using bilinear layers for increased interpretability and analytical tractability.
A fork of the classic Tensorflow Playground geared towards understanding Transformer training. Seeing is believing!
A repo that aids in basic research of neural networks via small-scale models. Run a GPU-enabled Docker env, easily define experiments, save model state, inspect weights, and plot results. A lightweight Weights and Biases.
I built a system that explores cognition via Francois Chollet's ARC (see link), combining the principles of program synthesis and minimization of algorithmic complexity.
Quickly build interactive dashboards in Python! Bento is an abstraction layer and templating engine on top of Plotly Dash, allowing a user to insert prefab widgets instead of building all from scratch.
- Plotly Dash