What the intelligence does

Eleven minutes, honestly estimated

Between a radar frame and a family moving the car upstairs sits a small stack of models: tracking cells, reading depths off kerbs, suppressing its own false alarms, routing around water — each trained on one council's ground and answerable to one council's threshold.

Six capabilities

Radar in. Sentence out. Act on.

Cell-tracking fusion designed

Nowcast the cell, not the state

MetMalaysia radar composites are tracked per convective cell and projected onto this council's catchments: where the rain that is falling now lands in eleven minutes, and what that nullah did the last six times it rained like this.

Depth-from-reference model

A photo of a kerb is a gauge

Vision models estimate water depth from crowd reports against kerbs, doorframes, parked cars and drain grates — the same measurement trick flood scientists use with reference objects, sized for a phone camera at dusk.

Multi-signal corroboration

The bell must stay believable

Every candidate alert passes a corroboration gate: radar trend, telemetry, neighbouring sightings, historical behaviour of that exact stretch of road. False alarms are the failure mode that kills civic alert systems — suppression is a first-class model, not a threshold.

State-machine routing, offline-safe

Route around the water, not the map

Evacuee and responder routing consumes live street states — flooded, gated, capacity-full — and degrades to the offline engine when towers die. The bus gets the forbidden-turn list the generic navigation app will learn about from tomorrow's news.

Per-instance models only

Each council learns its own ground

Catchment models train on one instance's history — its drains, its illegal dumps, its particular way Jalan Pasar gets wrong. No council's learning ever touches another's: the air-gap doctrine applies to gradients as well as rows.

Plain-language layer designed

Say it in one sentence people act on

Alert copy generation turns a probabilistic street-state forecast into 'left lane floods at the PJS 8 underpass around 5:10 — take the upper route' in Bahasa and English. An alert that needs a meteorology degree is not an alert.

Refusals as features

How far the models are not allowed to go

Probability, published

Every nowcast ships with its confidence and its evidence mix — radar-confirmed vs crowd-hunch is a visible label, never a blended secret. The council decides what 60% means on its streets.

Thresholds belong to the signatory

Models estimate; elected officials and their engineers set the trigger levels for alerts, closures and evacuations. Vendor-set thresholds are how alert systems get ignored or sued.

Never louder than the evidence

The product refuses phrasing the data can't carry — no 'historic deluge', no certainty the cell didn't earn. Honesty under uncertainty is the property that makes people move when the bell finally rings.

Every capability here exists because a December taught Malaysia what the alert layer was missing.

Why this shape — the evidence floor→

Ready to build something intelligent?

Let's engineer your next product — from idea to launch.