Thomas Lin

If what you want is a flight path, a transmission corridor, or a river, a bounding box is mostly bytes you paid for and threw away. This plans a request set for the actual geometry, coalesces it into HTTP byte ranges, and decodes DEFLATE partially from indexed mid-stream checkpoints.

Against tuned GDAL on live AWS Sentinel-2: 51% fewer bytes over the wire, 89% with precomputed indexes, 2.7× faster single-threaded and 15× in parallel — and byte-identical output across more than 15,000 reads, which was the part worth proving.

Stage is routed through the terrain with a HAND formulation in a numba core, then depth becomes dollars through USACE/HAZUS and JRC depth–damage curves, with a Monte Carlo uncertainty budget carried all the way to the final figure.

Validated across 16 watersheds nationally against 1,287 surveyed high-water marks. Terrain routing runs sub-second; fetching the DEM is what actually costs time, which is why the service is synchronous with no job queue. Self-hosted on a Mac mini behind a Cloudflare tunnel.

A physics-informed approach: a two-stage residual diffusion model downscales ERA5 temperature to HRRR resolution over central Oklahoma, trained with a loss weighted toward the tails of the distribution (λ=5, k=5). Extreme-event CRPS, which scores the whole predicted distribution against what was observed, improves 7%, from 0.858 to 0.799, for a deliberate cost in overall RMSE (1.31 to 1.48 K).