One US map engine. Any dataset, as a layer.

Born from a real product, not designed on a whiteboard

Allergy Locator shipped first: a full, validated US allergy-severity map with real ground-truth data, county-level gradients, and day-by-day forecasting. Once a second real dataset (healthcare access — nearest hospital by drive time) got built the same way inside that project, the shared shape between the two became clear enough to actually generalize — rather than guessing at an abstraction from a single case.

The full reasoning behind the split lives in Allergy Locator's own writeup.

Add a dataset by implementing one small contract

Every dataset — no matter how different its underlying data looks — implements the exact same shape: a fixed list of layers, an optional list of real years it has history for, and a lookup that returns a 0–100 value on a shared "higher = more concerning" scale, plus a one-line human-readable detail.

// lib/datasets/types.ts
interface Dataset {
  id: string;
  label: string;
  layers: DatasetLayer[];
  supportsTime: boolean;
  availableYears?: number[];  // real years only, e.g. crime's [2020..2024]
  getValue(cityId, layerId, context?): {
    value: number;   // 0-100, higher = more concerning
    detail: string;  // "72/100, high" or "Bellevue Hospital, 4 min"
  } | null;         // null = a real gap, never a fabricated value
}

That's it. Implement getValue, register the dataset once, and the exact same MultiLayerMap/AddLayerPanel components render it — continuous gradient, click-to-select detail panel, legend, a year control if it has real history, all of it — with zero new rendering code. Any number of layers from any number of datasets can be active on the map at once, each stacked as its own gradient (Allergy Locator's proven stacked-opacity pattern, generalized across datasets instead of just allergens).

Forty-one genuinely different real implementations

All free, all keyless, across the full 512-city spine — every one of them ships its own methodology doc naming its real sourcing and honest gaps. A few that show the sourcing rigor best:

Every dataset, every real source — named

In the spirit of the "never fabricate, always disclose" rule this whole project runs on: all 41, with exactly where their numbers come from.

Every row above links to the real, live, public source it was fetched from — no scraping behind a login, no purchased panel data. Full sourcing detail, real coverage numbers, and known gaps for each one live in this repo's own data/*-methodology.md files.

The same real data, from more angles

Roadmap