Forecasting that still works after handover
A demand model trained on your history, served behind an API, retrained on a schedule, and explained well enough that a planner will actually use it.
- Year
- 2026
- Duration
- 8–12 weeks
- Team
- 2 engineers
- Sector
- Machine learning / Operations
The problem
Planning runs on spreadsheets and instinct. Overstock ties up cash, understock loses sales, and nobody can explain after the fact why either happened. Most forecasting projects then fail twice over: the model is never trusted, and it silently decays once the person who built it moves on.
What we did
We clean the history, engineer features from it, and test several model families against a held-out period rather than against the training data. The winner is served behind an API and surfaced in a dashboard where planners compare forecast against actual and see which factors drove each prediction. Retraining runs on a schedule with metrics tracked across runs, and an alert fires if accuracy drifts past an agreed bound.
What you get
A forecast your planners can interrogate rather than obey, a service your engineers can call, and a retraining job that keeps it honest without us in the room.
Tell us what is breaking.
Send the problem, not a polished brief. An engineer reads every enquiry and replies within two working days with a real technical opinion — including when the honest answer is that you do not need us.
- Reply from an engineer, not a salesperson
- Fixed scope and price before any build starts
- Your code, your cloud, your repo — from day one