Case study · Personal project · 2026
ForeRoute — weather-aware safer routes
A Boston driving companion that pairs a rule-based weather score with a crash-history verdict on every stretch of a route.
- Role
- Personal project, built end to end by Daniel
- Building
- Cinema
- Year
- 2026
Problem
Drivers need safer routes that account for weather right now and the crash history of each stretch — without confusing raw model probabilities for crash odds.
Daniel's role
Daniel's personal project (2026), built end to end by him. ForeRoute suggests safer driving routes around Boston, taking weather into account. Each stretch shows two side-by-side badges: a rule-based 0–100 Weather right now score, and a Crash history verdict from a gradient-boosting model.
System diagram
Stack
- Next.js
- Mapbox
- OpenWeather
- MLflow
- Gradient Boosting
- SHAP
- Evidently
- Google Cloud Run
Outcome
Gradient Boosting chosen over Decision Tree and MLP by ROC-AUC: 0.65 on 17,808 held-out rows (including synthetic non-crash samples). Trained on 22,260 Boston Vision Zero crashes (2018–2025). Served on Google Cloud Run with SHAP explanations and Evidently drift monitoring.
Pair a deterministic weather scorer with a plain-language crash-history verdict so the product stays useful when the model is down — and flag stretches where the two disagree.
Field notes (measurements)
ROC-AUC
0.65 on 17,808 held-out rows (incl. synthetic non-crash)
Crashes
22,260 · 2018–2025