The intelligence of the road
Nine ideas, each with its honest state — and the three refusals above them. On a navigation app, the question is never what the model can guess, but what the system is allowed to say it knows.
Three refusals, before any feature
Nine ideas, honestly labelled
Live means shipped and device-verified; in design means specified against real data and owed a build; concept means it is a direction we are naming early.
The honesty gate — grey means unknown
A segment gets a traffic colour only when enough recent samples exist (five, inside forty-five minutes). Below that the road renders grey. This is the load-bearing rule of the whole product: the density trap that makes every small Malaysian app look useless is met with truth, not paint. Correct — and it must never be weakened by fabrication.
Provenance you can see
Every fused value carries its dominant source, and the app badges it: Community, Live traffic, Typical. You always know whether a colour is your fellow drivers' truth, a subscribed feed's, or a learned prior's. As segments flip to Community, the badge becomes an earned trophy — the network-effect proof, but truthful.
The self-retiring scaffold
Cold start solved without renting the map forever: licensed feeds and learned priors carry day-one colour, but only the crowd's weight grows with use. Where the community becomes self-sufficient, feed polling stops — cost shrinks exactly where trust was earned. The dependency is designed to be outgrown, road by road.
Kinetic crowd sensing
The road state inferred from the collective motion of anonymous phones — physics as the likelihood, not thresholds as the guess. Car-following models say whose brake is surprising; kinematic-wave theory says whether those brakes form a line moving upstream at a shock-plausible speed. A cluster that phases with a mapped traffic light is benign; one mid-block, aperiodic, is a story.
Software-defined positioning
The reference device has three sensors and no gyroscope — so the smart GPS is built in mathematics, not bought in hardware: a constant-velocity Kalman filter coasts through fix dropouts, map-matching snaps to the road you're actually on, and during re-planning the camera moves with the truth instead of dragging the truth to fit the map.
Asking, sparingly
When the app prompts “cars braking ahead — see it?”, scenario-aware rules decide: quiet when the slowdown is just your announced turn or a red light, progressively longer gaps between asks, three per trip, maximum. Every verdict is a ground-truth label — the crowd's own answers calibrate the system that asked. No annotation lab required.
Bus-first, from open data
Malaysia already streams live bus positions as open data. The feed gives positions, not ETAs — so the engine computes arrival by projecting each vehicle onto its schedule, clamps the known-bad fixes, and degrades unstable rail to schedule-only. Honest ingestion of what exists, rather than pretending to sense what doesn't.
The fleet that warms the road
Courier runs are sequencing problems first — time windows, capacity, a depot pinned — and every fleet ping that navigates a route also feeds the crowd substrate. The two-app fracture every dispatch SaaS accepts (sequence here, navigate on someone else's map) is dissolved because both halves are ours.
The leading forecast
Because a disturbance travels upstream at a measurable speed, the same physics that detects a hazard can predict when it reaches you — an advisory minutes before the queue comes into view. But a prior must never steer routing on speculation: the forecast only speaks once the crowd confirms it.
None of this is decoration — every rule traces to a measured failure, a licence term or a published piece of traffic physics. The floor is on the Evidence page.
See the research floor→Ready to build something intelligent?
Let's engineer your next product — from idea to launch.