Never guess what to order again.
What to order, how much, and why — learned from what the business has actually used, not from a fixed minimum somebody set two years ago and nobody has revisited.
| Corned beef 150g | sells 96/day | 4 days | order |
| Milk 1L × 12 | sells 42/day | 9 days | order |
| Cooking oil 1L | sells 31/day | 18 days | |
| Rice 25kg | sells 14/day | 29 days | |
| Canned tuna | seasonal — Lent | 44 days | hold |
| Overstocked | cash sitting on shelves | 38 items | ₱1.8M |
Machine-learned demand forecasting with reorder points, days of cover, seasonality and an accuracy figure that is measured rather than claimed.
What it actually does.
Trained on your own history, not a rule of thumb
Gradient-boosted models on your actual usage, backtested against what really happened — the same family of model large retailers use for exactly this. The accuracy figure is computed from held-back data, which is the only kind worth printing.
- Day-of-week and seasonal patterns learned rather than configured.
- Days of cover, not just quantity on hand — an item with plenty on hand and two days of cover is the urgent one.
- The last day you can order and still have it arrive in time.
It says why
Every proposed quantity carries its reason: the usage it is based on, the cover it leaves, the lead time it assumes. A number with no reason behind it gets overridden until nobody uses the system.
- Override where you know something the model does not — the override and reason are kept.
- Overrides are measured too, so the model and the buyer can both be judged.
- Move stock instead of buying it, where another site is holding.
Cover, not quantity. Eight hundred kilos means nothing until you know how fast it leaves.
Nobody has to be the expert
A new supervisor orders as well as your best one in their first week. The knowledge stops living in one person’s head, which is the risk nobody prices until that person leaves.
- What an order commits you to, before you place it.
- Forecast accuracy tracked over time, per item and per site.
- Seasonality by day of week, so a Monday order is not a Friday order.
The ones people ask about this.
How much history does it need?
It will produce something from a few weeks and it gets meaningfully better after a season. Where there is too little history to be honest, it says so instead of guessing.
What model is it?
Gradient-boosted trees — LightGBM — backtested against held-back periods. The accuracy figure on the screen is from that backtest, not from the training data.
Can buyers override it?
Yes, and they should when they know something it does not. Both the override and the reason are recorded, so the two can be compared later.
Does it handle promotions?
Promotional periods are known to the model, so a spike during one is not mistaken for a new baseline.
See it on your own figures.
Bring one ordinary day from your business and we will run it through in front of you, on your own items and your own prices.