Interface stills
The delivered system
Screens from the delivered system, deliberately obscured. We publish structure and density, never a client's data or naming — the same protection your project would get.
A machine-learning decision platform that predicts shipping lead times from historical order data, simulates the effect of reassigning production between factories, and ranks the resulting options by combined time saving and margin impact.
Orders were assigned to factories largely by habit and geography. Nobody could answer, with evidence, whether shipping a particular product from a different site would be faster, cheaper, or both — and there was no way to test the idea short of trying it on real customers.
Thousands of completed orders already contained that answer. They just sat in exports nobody could analyse.
We trained a Random Forest regressor on the full order history — product, factory, destination region, quantity and seasonality — to predict shipping lead time for any combination. It was evaluated against Linear Regression and Gradient Boosting baselines and selected on held-out performance.
A what-if layer that takes a proposed reassignment — move a share of one factory's volume to another — and projects the resulting blended lead time, blended margin and capacity utilisation before anything changes in the real world.
Rather than leaving the user to explore combinations manually, the platform evaluates every viable factory pairing and ranks them by a priority score combining lead time saved with margin percentage, so the highest-impact moves surface first. Options that would breach a factory's safe capacity ceiling are penalised automatically.
Delivered as a four-tab decision workspace:
Built with Python, scikit-learn, Streamlit and Plotly, with a dark and light presentation mode, deployed to a hosted environment with the source in the client's own repository.
Factory allocation moved from an opinion to a modelled decision. Instead of debating whether a reassignment would help, the team could see the predicted lead time change, the margin effect and the capacity consequence in the same view — and had a ranked shortlist of the moves most worth making.
The operations simulator on our live demo dashboard reproduces the same interaction model against sample data. Move volume between factories and watch the blended lead time, margin and utilisation respond.
Screens from the delivered system, deliberately obscured. We publish structure and density, never a client's data or naming — the same protection your project would get.
The same interaction model, live on this site, against sample distribution data.