The alert is arithmetic, not a model
AnBanjir uses machine learning at its edges and refuses it at its core. A warning fires on a council-set threshold — a line a YDP can defend — while models only extrapolate, corroborate and propose routes, under guardrails that keep a model from ever ordering an evacuation, naming a resident, or inventing a danger.
The threshold decision itself
This is the deliberate non-model at the centre. An alert fires because a measured or nowcast level crossed a line the council engineer drew — arithmetic, not a prediction. No model votes on whether to wake Lorong Sepuh tonight, because a probabilistic evacuation would be indefensible the first time a YDP is asked to stand behind it.
Radar-cell extrapolation
A model carries the next hour of a MetMalaysia cell forward across the catchment, so the console sees a street rising before a gauge reaches it. It proposes a level and a lead time; the council's threshold decides whether that becomes a bell. Forecast, never command.
Waterline corroboration
Matching a crowd photo against gauge telemetry, hazard layers and neighbouring sightings to score how much a cell should move. It raises or lowers confidence for the threshold to act on; it does not delete a report silently, surface a reporter, or decide an alert on its own.
Evacuee route planning
Choosing a path around flooded streets to a shelter with capacity, recomputed as the water moves. Genuinely useful, genuinely consequential — so it is planned to run entirely inside a single council's partition, on street state, never on a cross-council or cross-agency view of people.
Three lawful inputs. One street, eleven minutes out.
Federal radar, the council’s own gauges and a corroborated crowd waterline fuse in a per-catchment nowcast that runs on thresholds the council engineer set. It doesn’t answer “danger, somewhere” — it answers which road, this hour, and fires the bell scoped to that road when its level crosses the danger line.
This is an illustrative model, not a running nowcast. The figures are a shape, not a measurement — no live council instance exists yet. The rule it draws is the promise: resolution down to the street, the hour and the direction, with the alert scoped to the road that is actually rising. AnBanjir is designed, not yet running.
The guardrails
A model never orders an evacuation
Nothing in the ML layer issues a command, names a target, or closes a road on its own authority. The furthest a model reaches is proposing a level, a lead time, or a route for a human council officer to accept. The YDP and the engineer, who answer for the underpass, do the rest.
A model never names a resident
Every model reads inputs already inside the vault and writes only to the corroborated, aggregated side. No model is given, or produces, an output that identifies who reported what — the same no-join discipline that constrains the whole product constrains the training data too.
Ambiguity stays a silence
Where a photo can't be corroborated, a cell is thin, or a deluge outruns the model, the answer is "unconfirmed" — not a confident bell. AnBanjir would rather wake nobody than let a model cry wolf and spend the trust the alert is made of.
The honest summary: the model is not the promise — the offline threshold is. AnBanjir is interesting because of what its arithmetic makes defensible and what its partition makes impossible, not because of any model it runs. Where learning helps, it helps at the edges and under a rule it cannot cross; where it would be tempting to let a model order a road cleared or a street woken, we simply do not.
Follow the thread deeper
Warn the street. Route the rescue. Keep the map alive when the tower dies.
AnBanjir is designed, and it is built to be adopted one council at a time — a single PBT owns its catchment, its thresholds, and its console; nothing joins across councils. The opening offer is a reference deployment: one flood-prone catchment, one control room, one offline-capable state engine. Council enquiries are welcome today.