DataTwinLabsTraffic forecasting with MLOps
Deep-learning forecasts of speed and delay 24 hours ahead, trained, versioned and redeployed automatically as traffic patterns change.
01 — The problem
The city could see congestion as it happened, but not what was coming. Planners needed a forecast they could trust, and a way to know when it stopped being accurate.
02 — What I built
- ✦Feature pipeline combining live traffic, calendar and weather signals
- ✦PyTorch sequence models benchmarked against statistical baselines
- ✦MLflow experiment tracking and a model registry with staged promotion
- ✦Drift monitoring that triggers retraining through CI/CD
03 — What changed
A forecast shown next to live conditions in the twin, with every prediction traceable to the model version and data that produced it.
04 — Tools
PythonPyTorchMLflowFastAPIDockerGitHub ActionsTomTom API
Next project
Real-time transaction alerts for a bank