An LLM pipeline that turns disaster-risk-reduction reports into georeferenced event records with a fixed schema and controlled hazard vocabulary. Applied to 10,000 PreventionWeb documents, it produced 3,572 records across 24 hazard types.

A robust multi-fidelity Gaussian process that combines sparse, high-quality reference monitors with dense but noisy citizen-science sensors, using a Huber loss to bound the influence of contaminated data. Submitted to the Annals of Applied Statistics.