01 / H-Net20 min

What a learned tokenizer actually learns

A tokenizer-free model trained on SMILES, under four different objectives.

02 / IceCube10 min

Tracking neutrinos in a cubic kilometre of ice

A custom transformer, with attention biased by distance in space and time.

03 / Ribonanza10 min

One matrix multiply for RNA folding

A custom transformer, with base-pair probabilities multiplied into its attention.

04 / Contrails9 min

Segment Anything, stripped down for contrails

A U-Net decoder on the Segment Anything image encoder, with no prompts.

05 / Fast or Slow8 min

Predicting TPU runtimes on 10,000-node graphs

A graph transformer that orders configurations from fastest to slowest without running them.

I like dropping into unfamiliar problems, building models from scratch, and figuring out what actually works.

My path into machine learning started with a PhD in structural biology and bioinformatics. Today I work in ML, and outside of work I keep running independent experiments whenever a question won't leave me alone.

A lot of how I think about modelling comes from competitive machine learning. I love it because there's nowhere to hide from a hard benchmark, and every few months you get to jump cold into a completely new domain: neutrinos under the South Pole, RNA folding, satellite contrails, gravitational waves, compiler graphs. Working alongside great teammates, that turned into 10 Kaggle gold medals (peaking at 35th out of 200,000+ competitors) and a handful of peer-reviewed papers.

This blog is where I write up the full trail, including the dead ends that took weeks to figure out.

Publications

Competitive ML

Get in touch

I'm always happy to talk about machine learning, scientific AI, competitive ML, or any stubborn problem you're working on. If you're building something ambitious and think I might be a good fit, feel free to reach out.