Machine Learning Research Engineer

The Deep Learning team at WindBorne builds the best weather models in the world. We design and train a foundational model that ingests atmospheric observations and can produce global forecasts at high frequency, then we push it into production and prove it beats the national weather agencies. It's a small team that owns the full stack: architecture, datasets, data assimilation, training, and evaluation. In addition to research, we get our hands dirty with the hardware and optimize our systems to the max. We need more people who can do all of these things.

What you'd work on:

Own experiments end-to-end: architecture changes, loss function design, training runs, and rigorous evaluation against operational baselines. The research direction changes when results come in, and you drive it.

Understand real-world data, then model it: dig into messy real-world datasets like severe weather events and energy market variables, then extend our foundational model to predict them.

Evaluation & scientific rigor: weather forecasting is not as simple as maximizing a Kaggle score. You'd help make sure we're actually getting better, not just overfitting to metrics.

You'd be a good fit if you:

Have trained large models from scratch. You understand distributed training, gradient dynamics, and what it feels like when a run is going sideways at step 30k
Care about the science, not just the engineering. You've read papers to understand a problem domain, not just to replicate architectures
Are comfortable getting your hands dirty with hardware and the failure modes of real operational systems, not just clean research problems
Are comfortable with messy, real-world data that doesn't come in neat CSV files: satellite radiances, irregularly-spaced observations, multi-source fusion
Can move fast without a product spec. The research direction changes when results come in, and you're energized by that
Nice to have: Atmospheric science or physics background, experience with weather/climate data, PyTorch distributed internals, data assimilation or inverse problems.