xtream

Three of the 10 published projects describe machine-learning or optimization systems that reached production with numbers attached: hourly power-load forecasting for an unnamed Italian utility that cut relative error by 24% (residential) and 26% (industrial) and became the utility's official tool, demand forecasting for an unnamed food-and-beverage group, retrained monthly on Azure ML, that reduced planning effort by 25%, and WeRoad's tour planner, where 84% of departures now fall on a date chosen by the solver.

The stack is named each time (scikit-learn, PyTorch, a Luigi pipeline deployed on-premises, Azure ML with a Power BI front end) and the DeepPsyMe research write-up states plainly that its prototypes are not ready for production, which is the kind of disclosure most portfolios omit.

The two strongest ML deployments are anonymized, so the speakable reference is WeRoad, and the AI Solutions page rules out dashboards and image processing outright—useful scoping if your brief is forecasting, a mismatch if it is computer vision.

The weaker fit is buyers who want a large bench or a generative-AI showcase—the team page lists 20 people and every published AI study is predictive modeling or optimization, with no LLM deployment among them.