Where each core GIS concept is load-bearing in this system — not where it could be mentioned, but where removing it would break something.
| Concept | Where it is load-bearing |
|---|---|
| Datum | WGS84 throughout. The Survey of India’s legacy Everest 1830 datum differs by 200–300 m in this region — about 20 level-3 cells — so any toposheet coordinate must be datum-shifted before it enters the system. See accuracy. |
| Coordinate systems | Two, never mixed: EPSG:4326 for input and interchange, EPSG:32644 for all arithmetic. Degrees are not a metric unit, so no cell maths happens in geographic coordinates. |
| Projection | UTM zone 44N, chosen for conformality and metric units over a 50 × 60 km extent. Web Mercator rejected because its 2.6 % scale error at 13°N would make cells non-square on the ground. |
| Scale & distortion | The point scale factor is computed live and shown on every address. The distinction between absolute distortion (~400 ppm, dominated by UTM's k₀) and differential distortion (~160 ppm, the number that decides equal-area) is made explicit. |
| Vector data | The CMA extent, ward and landmark geometry, and every cell footprint — cells are emitted as true polygons whose corners are projected individually, which is why they tile without seams. |
| Raster / tessellation | The nested grid is a raster-like regular tessellation. Its row-major indexing, and the conversion to the north-first raster convention, are part of the cell model. |
| Topology | Containment is exact prefix comparison; adjacency is ±1 in a mixed-radix global lattice. Both are exposed in the UI and in /api/neighbours. |
| Data quality & accuracy | Cell resolution versus GNSS positional accuracy, quantisation RMSE, and the datum trap — all quantified rather than asserted. |
| Data sources | OpenStreetMap, Bhuvan/NRSC, Survey of India and Census ward data — roles and licences on the sources page. |
| Input / output | GNSS input via the browser Geolocation API; web map output via MapLibre or Google Maps with a shared deck.gl overlay; machine output via a documented JSON API returning GeoJSON footprints. |
| Generalisation | Progressive disclosure is cartographic generalisation applied to identifiers: truncating an address is exactly zooming out a level. |
| Spatial analysis | The error-envelope analysis is a containment query over the hierarchy; the confusion set is a nearest-neighbour query in embedding space projected onto geographic distance. |
Where the AI component sits
The word layout is produced by a sentence-embedding model, a Kohonen self-organising map and a Hungarian optimal assignment, with the result measured against a random-layout baseline. That is described in full on arranging the words, and it is the part of the project that is genuinely a machine-learning result rather than a piece of engineering.