Work

Software

Here are some open source software libraries that I’ve made substantial contributions to.

PyMC3 and PyMC4

PyMC3 is a popular Python framework for Bayesian modeling and probabilistic machine learning, focusing on Markov chain Monte Carlo (MCMC) and variational inference (VI) algorithms. I’m a member of the core development team, and contribute to the PyMC3 internals and documentation. I wrote a blog post on tips and tricks for Bayesian modelling using PyMC3.

Pyfolio and Alphalens

Pyfolio and Alphalens are Python libraries for risk analysis and performance attribution of financial portfolios, and alpha factor research for algorithmic trading. Both libraries are fully integrated into the Quantopian platform.

I developed the risk and performance attribution capabilities of Pyfolio (read more on my blog post here), and help maintain the library. I help develop new features, triage bug reports and troubleshoot issues for Alphalens.

Knead and Glaze

Knead and Glaze are command line tools and Python libraries for preprocessing, manipulating, rendering and visualizing font files and algorithmically-generated typefaces. Both libraries were written as internal tools for The Font Bakers, to support their research and development workflow. I developed and maintain both projects.

Datasets

cryptics.eigenfoo.xyz

cryptics.eigenfoo.xyz is a dataset of cryptic crossword clues, indicators and charades, collected from various blogs and publicly available digital archives.

The project scrapes several blogs and digital archives for cryptic crosswords. Out of these collected web pages, the clues, answers, clue numbers, blogger’s explanation and commentary, puzzle title and publication date are all parsed and extracted into a tabular dataset. The result is over half a million clues from cryptic crosswords over the past twelve years.

Projects

Here are some interesting machine learning and data science projects I’ve pursued.

Deep Learning for Algorithmic Type Design

As a research project pursued in my last year at The Cooper Union, we researched a class- and attribute- conditional generative adversarial network capable of producing vector graphics, with potential applications in algorithmic type design. The generative model was to produce closed shapes, with counter spaces, defined by a variable number of quadratic Bézier curves. At the time, no generative model with vector graphic output has previously appeared in literature.

Hate Speech on Reddit

As part of a project on Data Science for Social Good, I ran text clustering algorithms on well-known hateful and toxic subreddits, and collaborated with a cross-disciplinary team of artists, architects and engineers to present the findings at The Cooper Union 2018 End of Year Show. I also wrote a blog post on my results, and gave a talk on the data science that went into the project.