Online coding and Python classes in Edgbaston, Birmingham

The learner downloads OpenStreetMap data for a rectangle of south-west Birmingham that includes Edgbaston, and collects every mapped place that has a name and is tagged as a shop, amenity, office, leisure, tourism, craft or healthcare site: 4,842 places. Each name is lower-cased, stripped of punctuation and cut into overlapping three-letter pieces called shingles. Two names are compared by their Jaccard similarity: the shingles they share divided by all the shingles either has. Comparing every pair takes 11,720,061 comparisons and 23.5 seconds in plain Python, and finds 1,901 pairs with a similarity of 0.6 or more.

MinHash replaces each set of shingles with a short signature, here 128 numbers, built so that the chance two signatures agree in any one position equals the Jaccard similarity. Locality-sensitive hashing then chops each signature into bands, and only names that match exactly on at least one whole band become candidate pairs. More rows per band means stricter matching and fewer candidates.

With 32 bands of 4, the program examines 11,152 candidate pairs, about one in a thousand of the full set, and still finds every one of the 1,901 similar pairs; building the signatures took 4.7 seconds and the bucketing half a second. Push the setting further, to 16 bands of 8, and the candidates shrink to 1,820 but 14% of the true pairs slip through. That is the trade every user of the method has to make. Are the pairs really duplicates? Only some. Of the 1,901, just 288 are within 100 metres of each other, and 90 share the same type of place within 30 metres. Many of the rest look like branches sharing a brand name, or two features of one site such as an attraction and its car park. A similar name is evidence, not proof.

Understanding the shortcut, and its price, lets Edgbaston teenagers use large-scale AI tools with their eyes open, which is what learning Python is for. The longer argument is in why teenagers should still learn to code in 2026.