Building a City Digital Twin

The arc, in one breath

Part 1 writes down the mental model I began with - a city as entities on geography, flows between them, and senses that watch both - then records the six corrections the literature and the data forced on it: senses that report at speeds seven orders of magnitude apart, sensors that are a shadow of the state rather than a mirror of it, people who re-optimise rather than drift like particles, single runs that hide their own uncertainty, governance that is the controller of the loop rather than another entity inside it, and an output that should compare futures instead of forecasting one. Every correction pushes the same direction - toward a smaller, scoped twin. Part 2 puts that argument on trial against the record: a decade of government everything-twins that each converged to a scoped survivable core - a geometry platform, a flood demonstrator, a framework document - the scoped products that shipped and stayed shipped, and a US$2 billion foundation-model company running the same experiment today, whose independent validators have already drawn the envelope around its scoped part. It closes on the capability nobody has built: a spatial twin coupled to a council’s money. Part 3 stops arguing and goes to stand inside one of the twins the record vindicated: the NSW Spatial Digital Twin. On census night it takes the portal up on its “add your own data” button and pushes a census-night headcount down onto individual land parcels, a grain the Bureau of Statistics never publishes at, which turns out to be a lesson in what a scoped twin will and will not tell you. Part 4 stops reading and starts building: a scoped twin made by hand, with OpenStreetMap roads cropped by osmium and run through SUMO, put to one concrete question - how long the Victoria Park precinct takes to clear when 63,000 people leave the Brisbane 2032 stadium at once, and where it chokes. Thirteen real runs are baked into an interactive map, with the assumptions and the sanity checks set out beside them. Part 5 shows the much bigger twin built alongside it: Greater Newcastle, where six hundred thousand synthetic residents generated from the census travel by twelve modes on the real roads and timetables, drawn in maps and in the twin's own 3D viewer, and checked honestly against how the real city travels. Part 6 opens it up stage by stage: the network and footpaths, the mapped timetables, adaptive signals and level crossings, the synthetic population and their days, the simulation that lets them choose, how each run is checked, and what it takes to run one. Part 7 adds a second city on the same framework - Mumbai, 27 million synthetic residents, its trains read from the printed timetables - and the way both were built, with AI coding agents that kept the whole project in the repository. Part 8 is what Mumbai led to: the proposal we prepared for a state government in India, published in full with its slides playing, and an honest map of which parts run today, which are real ground under illustrative layers, and which are a direction of travel.