Test a route, inspect the prediction.
Cabrynt uses OSRM for a road-route baseline, then applies an ONNX residual model using route and quote-time context.
Test a Porto route
Choose a pickup and destination inside the supported area.
Live prediction
Select two points and calculate a route to see the baseline and model-corrected prediction.
Model performance
The deployed model was trained on 200,000 historical Porto trips and evaluated on a separate set of 5,000 unseen trips.
Average error
38.4% lower
Mean absolute error compared with direct OSRM.
Typical error
53.1% lower
Median absolute error across evaluated routes.
Higher-error routes
38.0% lower
P90 error, the level not exceeded by 90% of routes.
| Model | MAE | Median absolute error | P90 absolute error |
|---|---|---|---|
| Direct OSRM | 5.488 min | 3.530 min | 10.817 min |
| OSRM + ML correction | 3.383 min | 1.656 min | 6.702 min |
Prediction scope
- Porto routes only.
- OSRM route duration and distance provide the routing baseline.
- Quote-time calendar, weather, route, and historical-congestion features provide context.
- ONNX Runtime serves the residual model within the ASP.NET Core API.
Known behavior
The model corrects a road-route estimate instead of recreating routing from coordinates. Direct OSRM was more accurate for completed historical trips lasting up to five minutes, but actual duration is unknown when a quote is requested, so this cannot be a dependable live fallback rule.
The API falls back to OSRM only when routing context, weather, or model inference is unavailable.