Avalanche Risk Forecasting

Built a physics-informed machine-learning model that forecasts avalanche risk at 2–5 km resolution. Inputs include satellite imagery, snowpack depth and stability indicators, recent weather, and terrain — features that physical avalanche models traditionally use, now combined with ML to handle their interactions.

Test accuracy reached 83% on held-out data. Temperature gradients and water-content features emerged as the strongest predictors. The open-source web tool returns risk estimates for any user-selected map coordinates.

I'm a longtime snowboarder, and the project started as a way to bring data science to mountains I actually ride. Current work focuses on recall, false-negative analysis, and data-leakage checks before I'd trust it for real safety use.

3rd Place, Engineering — Golden Gate STEM Fair (ISEF-affiliated regional).