Four modules
that pay for
each other.
Most travel tools bolt detection, search and booking together and hope. ViaMe runs them as one feedback loop: the prediction model weights the routing graph, routing constrains what booking can attempt, booking outcomes retrain prediction, and the device throttles its own sensors to the live risk score.
It sees the cell
before the gate does.
A transformer-class model fuses four asynchronous feeds into a single disruption probability per itinerary leg, refreshed continuously. Because the feeds disagree in useful ways — radar leads the airline's own status board by minutes — the fused score crosses the alert threshold while there is still inventory to rebook into.
Validated against historical disruption sets at a minimum 85% accuracy on the binary "will this leg break" task.
| Architecture | Transformer-class, ~10⁸ parametersmulti-head attention over time-aligned multimodal tokens |
| Output | Per-leg disruption probability r ∈ [0,1]alert threshold τ = 0.75, tunable per traveller profile |
| Refresh | Continuous · up to 50 Hz on active legs |
| Accuracy | ≥ 85% on held-out historical disruption sets |
Every edge knows
its own risk.
Transport is modelled as a directed multigraph — airports, stations and kerbsides are nodes; flights, trains and ride-shares are parallel edges between them. Each edge carries time, cost and the live risk score from the prediction model, so a risk-weighted A* search naturally avoids the legs about to break.
Instead of one "best" path it returns an n-best diverse set, so the alternatives don't all fail together when the same storm closes them.
| Graph | Directed multigraphparallel modal edges between transfer nodes |
| Search | Risk-weighted A*transfer-aware admissible heuristic |
| Output | n-best diverse routesdecorrelated failure modes |
| Latency | < 500 ms5+ leg international itineraries |
The model rides
in your pocket.
A distilled copy of the prediction model runs on the phone, so the risk score keeps updating when connectivity drops over an ocean or in a tunnel. Getting a 10⁸-parameter teacher down to an ARM-friendly student is a three-step compression pipeline — and the same risk score then governs how often the device is allowed to wake its sensors.
Distil
A compact student network learns to mimic the full teacher's risk outputs, preserving the decision boundary that matters around τ.
Prune
Structured pruning removes attention heads and channels that don't move the score, shrinking the graph for mobile memory budgets.
Quantise
Weights drop to 8-bit integers for ARM NPU execution — sub-200 ms inference with negligible accuracy loss.
The full architecture,
in one document.
The technical whitepaper covers the fusion model, the routing heuristic, the Saga compensation logic and the on-device governor in depth.