Things I have built, described the way I would describe them out loud.
Ongoing, at APQX
Industrial R&D moves slowly because the expensive step is not the thinking, it is the waiting — for an experiment, for a simulation, for someone to notice a result. I build systems that act on real processes rather than just describe them, and that know the difference between a number they measured and a number they guessed.
2024
A constitutive model is a physicist's guess about how a material responds when you push on it, and there are dozens of competing guesses. Instead of picking one, this learns the whole family at once — so you can hand it a material it has never seen and get a usable model back rather than starting the argument from scratch.
2024
Good data about a fluid is expensive and slow; rough data is cheap and abundant. This learns from a lot of the rough and a little of the good, which is how people actually work. We characterized a commercial shampoo from a fraction of the measurements it would normally take.
2022 —
Rheology-informed neural networks. A network that fits data will happily fit nonsense; one that also has to satisfy the physics has far less room to be wrong. This was the tooling for that idea, and it is open source because the idea is more useful than the credit.