Skip to content
All projects

End-to-End AI ProductsRail infrastructure

Trainyard

Which trains to inspect this week, with a governed model behind the answer

What does it cost when a train fails in service?

A regional rail operator runs 40 electric units, and a compressor failure in service strands the train: about five failures a month, at about €15,000 each. A predictive model exists, but every cycle takes the ML team 22 to 44 hours of manual work, so it is retrained once a quarter and nobody knows when it starts to drift. Trainyard gives the planner a weekly list of the trains to inspect, with the signals behind each one, and runs the model cycle unattended. The model changes only with its owner's approval.

recall at seven days
0.90
inference under load, p95
389 ms
median model cycle
7.1 min

Fictional client, synthetic use

Gallery

Architecture of Trainyard per environment

1 / 33
1 / 33

Architecture of Trainyard per environment

Built with

  • Python
  • Docker
  • GitHub Actions
  • FastAPI
  • React
  • Locust
  • Microsoft Agent Framework
  • scikit-learn
  • MLflow
  • Bicep
  • Azure Machine Learning
  • Microsoft Foundry
  • Container Apps
  • Cosmos DB
  • Key Vault
  • Blob Storage
  • Application Insights

Next project

Juriscope

Cited answers on the Spanish Criminal Code, in the wording in force on the date