Deep learning anchored in physics
Computational science and ML engineering, where rigorous modeling comes to production.
Applied and professional work
Corn nitrogen response
Site-specific nitrogen dose-response curves, pairing canonical agronomic forms with statistical learning, plus an honest measure of how far each curve can be trusted. Separates site potential from response shape instead of black-box regression.
meandre
Differentiable hydrology. Process-based catchment models written so every flux and storage term is differentiable, enabling gradient-based calibration at scale, identifiable sensitivity analysis, and hybrid models kept physical by conservation laws.
The tangent/ suite
The work I'm proudest of: an open-source suite that makes scientific computing reproducible and install-free. Seven JavaScript modules, tested against the reference tools in R and Python. A data-science library, a Bayesian engine, and the note interface that ties them together.
Data-science toolkit for JavaScript: ordination, clustering, classical stats and ML. The ergonomics of scikit-learn and the tidyverse, tested against both.
DocumentationLocal-first computational notebooks that run JavaScript in the browser via WebAssembly. No server, works offline, your files stay on your machine.
note.tangent.toBayesian inference for JavaScript: probabilistic modeling and MCMC sampling in the browser, so a model can report how sure it is, not just what it predicts.
DocumentationNotes on getting modeling right
How I destroyed my own model
I built an automatic mineral prospector, then tried to break it with strict space and time holdouts. It didn't survive, and that is the point. A look at why some exploration data quietly teaches models the wrong thing.
Read the postMore posts on reproducibility, differentiable modeling, and the capyllary build, as the work ships.
Engineering rigor, natural-science context
I combine the precision of engineering and physics with the messy, nonlinear reality of agricultural, ecological, hydrological and geological systems. My through-line is reproducibility: models and tools that others can rerun, audit, and trust, whether a calibrated hydrological model for a government agency or a notebook that runs with nothing installed. Good science needs good tools, so I build both.
Full background on LinkedInBackground
- PhD, Civil Engineering
- Postdoctoral research, agriculture
- Former professor
- Modeling, government (current)
Let's talk.
A modeling problem, a collaboration, or a question about the tools. Get in touch.