Step 02 · Prediction Engine
In progressWhat will happen?
Forecast 8-week demand per SKU per depot, accounting for seasonality and weather.
A trained model per depot-SKU, with held-out accuracy reported on every forecast. Inputs: history, calendar (Ramadan, Eid, exam season), temperature, promo flags, cricket fixtures.
Held-out MAPE 7.4%
Terms on this page
- SKU
- One sellable product in one pack size. Each depot-SKU pair gets its own model.
- Held-out accuracy
- Accuracy measured on data the model never saw during training.
- MAPE
- Mean absolute percentage error: the average gap between forecast and actual demand. Lower is better.
How the Prediction Engine gets there
- 01Three inputs: results already surfaced by the Analytics Engine, the underlying database, and the user’s prompt.
- 02Our proprietary agentic structure reads the shape of the data (task type, target, features, volume) and composes a prediction specification.
- 03It selects candidate models, or an ensemble, from a curated catalog, then runs a real ML workflow: preparation, training, validation, held-out testing.
- 04The lifecycle stays prompt-driven: promote a model into production, wire it into a dashboard, schedule re-forecasts and alerts.
In progress. Under active development now: agentic AutoML with calibrated, explainable output for real business prediction targets.
Explore the Prediction EngineWhat changes
The relay today
Request routed to a tech lead, then a data team: gather, clean, build, validate, ship. Weeks or months pass.
With Theovya
A trained, validated model with honest metrics, while the question is still relevant.
The same three questions every distribution business asks. The difference is that here they are asked by the same person, on the same afternoon.