Method

Schools come from OpenStreetMap: 1,100 in Delhi, 823 with their grounds mapped, 242 identifiable as government schools. Buildings come from the Overture Maps release of September 2026, read straight from its public bucket on AWS. Only three of the release's 512 files overlap Delhi, so we select them from the release's STAC catalogue rather than scanning all of it: 1.9 million footprints in two minutes.

Trees come from Meta and WRI's global canopy height map, version 2, at about 1.2 m, also on AWS Open Data. Ground height comes from the Copernicus 30 m elevation model. Both are read as cloud-optimised GeoTIFFs, a few kilobytes per school, never downloaded whole. Weather is ERA5 reanalysis through Open-Meteo for Monday 27 April 2026, the hottest school day before Delhi's summer vacation: 42.2 °C at 2 pm, 14% humidity.

Open data has a gap. Only about 0.5% of Delhi's buildings carry a height or a floor count, so heights mostly come from a prior based on footprint size. At 11 am to 1 pm in late April the sun is 70 to 75 degrees up and building shade in a yard is small, so the ranking is far more sensitive to trees than to building heights. It still matters in the early morning.

SOLWEIG, from the UMEP project used in urban climate research, computes the full radiation budget of a person standing at each square metre: direct sun with shadows from buildings and tree crowns, diffuse sky light weighted by how much sky the spot can see, and longwave heat from sun-warmed ground and walls. Its output is mean radiant temperature. Combined with air temperature, humidity and wind it gives UTCI, the feels-like temperature behind the heat stress bands on every map here. We run UMEP's Rust implementation on the laptop GPU: a 250 m school takes about 4 seconds.

The ranking runs every school at 2 m for 9 am, 11 am and 1 pm: 766 schools in about ten minutes. Heat twins run at 1 m for every hour from 7 am to 6 pm.

SOLWEIG is too slow to follow a mouse. So the planner has two tiers. The preview recasts shadows for the edited yard in your browser and predicts the change from the physics run, not the absolute value: a per-hour linear model of how mean radiant temperature responds to losing direct sun, gaining a crown overhead, and having canopy nearby, fitted for each school against a second SOLWEIG run with trees scattered over the yard. Squares a design doesn't touch keep their physics value exactly. Where you plant, the preview can be off by several degrees at midday, and the page says by how much.

"Verify with physics" sends the design to a worker that runs SOLWEIG on the edited school and recomputes the child's day, about 10 seconds. The API and worker run in containers against an emulated AWS stack locally; the same images are built to run on AWS Lambda.

JOS-3 models the body as 17 segments, each with core, muscle, fat and skin, and simulates sweating, blood flow and shivering minute by minute. We scale it to a nine-year-old (1.32 m, 28 kg) and walk it through the Delhi government school morning shift: assembly outdoors at 7:30, class, recess at 10, class, and the walk home at 12:30. Outdoors it stands in the yard's median mean radiant temperature for that hour; classrooms are taken as slightly cooler than outdoor air, out of the sun.

JOS-3 is validated on adults, and its set points do not scale down: a child-sized body rests at about 38 °C core in the model. So we report the rise above the child's own resting state, which is what heat-strain studies compare, and water lost to sweat and breath. The model does not capture children's lower sweat rates. Read these as indicative.

Property tests check geometry that must always hold: a taller building never lights a cell that was in shade, and planting never adds sun anywhere. Differential tests compare independent implementations: our shadow caster agrees with SOLWEIG's own shadows on more than 97% of cells, the preview stays within 1 °C of SOLWEIG over a yard it has never seen, and the browser's TypeScript port reproduces the Python on a shared fixture. One test exists because SOLWEIG quietly rewrites the arrays it is given; the engine now passes it copies. The job queue is contract-tested against emulated DynamoDB and SQS, including a redelivered message running exactly once.

The 3D views draw the same grids SOLWEIG uses, one column per square metre: roofs from Overture and Copernicus heights, crowns from the canopy map, the ground coloured by the 1 pm map, rendered with three.js. Next is measured 3D. recon/ turns a 60 to 90 second phone video of the yard into camera positions and a dense point cloud with VGGT, then into heights on the 1 m grid. The same footage can train a Gaussian splat, rendered in the browser with Spark, World Labs' open-source splat renderer for three.js. TheirMarble world model points further still: whole worlds generated from a single photo. Anything generated stays a picture; the numbers stay physics.

SOLWEIG and UMEP (GPL-3.0), pythermalcomfort and JOS-3 (MIT), Overture Maps (ODbL), Meta and WRI canopy height (CC BY 4.0), Copernicus DEM, OpenStreetMap (ODbL), ERA5 via Open-Meteo, numba, DuckDB, rasterio, GeoPandas, FastAPI, Astro and React. Neem itself is GPL-3.0. The source is on GitHub.