Scientific software

Research tools for probabilistic machine learning and representation learning in biology.

I build tools that make scientific methods usable beyond a single analysis. My work emphasizes explicit assumptions, reusable interfaces, and documentation that connects the implementation to the underlying model.

BarBay.jl

Bayesian estimation of relative fitness from high-throughput pooled competition assays. I developed the method and authored the Julia package.

DNA barcode counts → Bayesian inference → relative fitness with uncertainty

  • Variational inference for analyzing many lineages.
  • Models for multiple environments and hierarchical analysis of replicates and genotype-linked barcodes.
  • A companion research website with methods, analyses, and figure-generation code.

Documentation · Source code · Worked research analysis · Research overview

AutoEncoderToolkit.jl

A Julia package for training variational autoencoders and their extensions. I authored the package, which is published in the Journal of Open Source Software.

High-dimensional data → autoencoder training → latent representations and geometric analysis

  • Built on Flux.jl, with interfaces for multiple autoencoder variants.
  • Utilities for analyzing the geometry of learned latent spaces.
  • Documentation connecting model training with research use.

Documentation · Source code · Software paper · Related research

SCRIBE

A framework I am developing for probabilistic modeling of single-cell RNA sequencing data, using Bayesian inference to account explicitly for measurement noise and uncertainty.

Single-cell RNA counts → probabilistic modeling → biological estimates with uncertainty

  • Scalable inference implemented in Python, JAX, and NumPyro.
  • GPU acceleration, variational inference, and MCMC.
  • Models addressing technical variation, including variable capture efficiency.
  • Integration with the AnnData ecosystem and tools for posterior analysis.

Publication in preparation. SCRIBE is under active development. The source code is private; this overview describes its capabilities at a high level.

Research overview · Contact me

Contributions to multi-agent learning

As a visiting researcher at SoftMax AI in May–June 2025, I contributed to the open-source Metta multi-agent reinforcement learning platform. I designed exploration curricula and reward structures, drawing on biological models to formulate learning objectives and evaluation metrics.

This work extended my research into how learning environments and objectives shape agent behavior. More about my background.