Who made this, and how
This page says who made the study, how it was built, the decisions that shaped it and how its results were checked.
My contribution
Role. Concept and brief, research questions, choice of cities and scenarios, design direction, and review of the results.
Built with Claude Code, an AI coding assistant, which wrote the pipeline and the site under my direction.
Dates. October 2026 (OpenStreetMap snapshot 5 October 2026; first public release 14 October 2026).
Source code. github.com/BLTC-520/sgjb-15-minute-city
How it was made
The study was designed around one sentence from the case study: a reader should feel the difference between a circle and a walk. The work ran in these stages.
- Brief and questions. How far everyday essentials are from home on foot, in two cities that share a strait, and how future mobility could change that.
- Open data, with checksums. Each source is recorded with its SHA-256 checksum and retrieval time in
data/provenance.json. - Street network and population model. A walkable network from OpenStreetMap, and census residents placed in hexagons.
- Accessibility for every inhabited hexagon. A walking distance to each of six essentials, and the share with all six in reach at each walking speed and time budget.
- Validation. Against official data and a second places source: bus stops and rail stations, and OpenStreetMap places against Overture Places.
- Sensitivity analysis. A 72-run analysis over walking speed, time budget, definitions, network rules and population model.
- Scenarios. The RTS Link, a shared autonomous feeder, car parks converted into neighbourhood hubs, and one new footpath in a named Johor Bahru neighbourhood.
- The site. Story, Explorer and Methods, all drawn from the same published data.
- Release. GitHub Pages, under a custom domain.
- Readable without JavaScript. Prerendered pages and an
llms.txtguide, so search engines and AI assistants can read the study. - Revisions after a teacher's review. The review asked for a clearer account of my role, the decisions and the development process; for pages that stay readable when the interactive map cannot run; for the qualifications beside the headline figures; for wording that claims no more than is measured; and for one detailed intervention in a named neighbourhood.
Three decisions
Three choices shaped the results most. For each, the card gives the decision, the alternatives considered, why this one was chosen and what it costs.
Measure walks on the street network, not circles
Distances are walked on the OpenStreetMap street and footpath network. Homes and destinations join it at the nearest point on the nearest walkable edge, not the nearest node. The off-network leg counts at 1.3 × its straight-line length, and nothing joins from more than 250 m. One multi-source Dijkstra search per essential serves every walking speed.
Alternatives. Straight-line buffers (the 1.08 km “15-minute circle”), or snapping to the nearest node.
Why this one. Circles overstate access where streets are discontinuous. A 15-minute walk reaches a median of only 47% (Singapore) and 34% (Johor Bahru) of the residents living inside a home's circle.
Trade-off. Results depend on how completely OpenStreetMap maps footpaths and crossings. Slope, heat, shade and crossing waits are not modelled. A stricter network variant is part of the sensitivity analysis.
Place people with census totals and a settlement map
Census counts (Singapore subzones, Johor Bahru state constituencies) are spread over the hexagons in proportion to Meta/CIESIN HRSL settlement weights, with Kontur weights where HRSL sees no settlement. Totals therefore match the census exactly.
Alternatives. Spreading people evenly over each census unit, which puts residents in forests and industrial land; or using a gridded population layer directly, which does not match census totals.
Why this one. The census is authoritative for how many people there are; the settlement layer is better for where they live.
Trade-off. Placement within each unit is modelled. The Methods page reports how well each model matches census units, and the Kontur model is an alternative in the sensitivity analysis.
A static site that computes in the browser
The Python pipeline precomputes compact binary tables and a tiled street graph. The browser recomputes completeness for any assumption and runs the walk-shed searches itself. There is no server, database or API key. The site is hosted free on GitHub Pages and prerendered for readers without JavaScript.
Alternatives. A server with a spatial database and a routing engine.
Why this one. Nothing to keep running or pay for, the published numbers are frozen and reproducible, and it loads fast.
Trade-off. Readers download the data files and basemap tiles the pages need, tens of megabytes for the Explorer. Updating the analysis means re-running the pipeline. The interactive maps need WebGL, and without it the site falls back to still images and a table.
How the results were checked
- Unit tests of the shortest-path code: the Python kernels against SciPy's reference Dijkstra, and the TypeScript search against a naive O(n²) reference (pytest and vitest).
- Cross-language tests (vitest): the browser decodes the Python-written tables, routes on the Python-written street tiles, and mirrors the Python metrics.
- Audits against official data. Official bus stops with an OpenStreetMap stop within 30 m: 92% in Singapore (LTA DataMall) and 83% in Johor Bahru (myBAS). Official rail stations with one within 200 m: 94% in Singapore (LTA).
- A full re-run with Overture Places as the destination source. All six essentials within 15 minutes: 89.5% of Singapore residents and 23.0% of Johor Bahru residents, against 80.4% and 12.0% with OpenStreetMap.
- 72 sensitivity runs, with each finding labelled robust, conditional or fragile.
- End-to-end checks of the built site: the prerendered pages, the Markdown and llms.txt copies, the URLs and the agent tools.
Tools
Python (NumPy, numba, GeoPandas, pyosmium), TypeScript, React, MapLibre GL, deck.gl, KaTeX, Playwright, Claude Code.
Keep reading
Start with the Story, try the Explorer, or check every formula and source on the Methods page.