Method
The complete technical account — including the parts that argue against the design. Every number on these pages is computed from the same modules the map runs on, so nothing here can drift from the system it describes.
- Working CRS
- EPSG:32644 — UTM 44N
- Levels
- 3 (+1 optional)
- Finest cell
- 12.5 m square
- Addresses
- 1,92,00,000
- Vocabulary
- 920 words, 3 disjoint domains
- Layout model
- all-MiniLM-L6-v2
The model
How does a coordinate become three words?
Datum and CRS choices, the UTM 44N metric grid, the nested tessellation, and the encoding algorithm end to end — with the worked example for Marina Beach.
9 min
Arranging the words
Why is 'copper' next to 'bronze'?
The NLP pipeline: sentence embeddings, a Kohonen self-organising map over each level's lattice, and an optimal assignment — plus the measurements showing it worked.
8 min
What hierarchy costs
Isn't a wrong word dangerous here?
The honest analysis of the trade-off against what3words: what the vocabulary contract catches, what it cannot catch, and how far each class of mistake actually puts you.
7 min
Accuracy & data quality
How wrong can this be?
Projection distortion, cell resolution versus GNSS positional accuracy, quantisation error and its RMSE, and the datum trap that would put you 200 m out.
6 min
Against the alternatives
Why not just use what3words or Plus Codes?
A structured comparison with what3words, Open Location Code, geohash and MGRS across hierarchy, topology, error behaviour and interpretability.
6 min
Data & sources
Where does the geography come from?
OpenStreetMap, Bhuvan/NRSC, Survey of India and Census ward data — what each contributes, their licences, and the vector/raster/topology roles they play.
5 min
GIS concept map
Which GIS ideas does this exercise?
An explicit map from core GIS topics — datum, projection, vector, raster, topology, data quality, input/output — to where each one is load-bearing in this system.
5 min