Probabilistic forecasts at SKU-location level, retrained weekly, with the error bands your planners actually read. Built on your data, deployed into the process you already run, and handed over when it is working.
One planner, one workbook, last year plus a gut feel. It works until the planner is on holiday or the category changes.
A single number per SKU hides the uncertainty. Planners quietly add their own buffer, and the buffer becomes the forecast.
A data science team built something in a notebook. It scored well and nobody used it, because it never fit the weekly planning rhythm.
Every SKU-location gets a P10 to P90 range, not one number. Safety stock and service levels follow directly from the band.
Weekly retraining on your latest sales, promotions and stock-outs. The model tracks the business instead of decaying.
Launches are forecast from catalogue attributes and comparable items, so a new SKU is not a blank row for its first season.
Outputs land in the tool your planners use, whether that is the ERP, a planning system or a governed sheet. No new login to remember.
Two weeks. We map how the forecast is produced today, measure its error honestly, and find the categories where a better model pays back first.
Several model families tested against your history, judged on the metric your business cares about. The one your team trusts wins, not the one with the best backtest.
Weekly retraining runs on its own. Your team gets the KPI board, the SOPs, and a direct line for changes through the client portal.
Forecast error (MAPE) on fresh categories at a grocery retailer
Launch-week stock-out rate after a cold-start model for new SKUs
Retraining cadence, with drift monitored between runs
Results from past engagements; clients undisclosed. Your numbers come out of the diagnostic.
Two to three years of sales history at the SKU-location level, current stock positions, and whatever you have on promotions and price. Messy is fine. Cleaning it is part of the work.
No. It replaces the part of their week spent producing a baseline by hand. They spend that time on exceptions, promotions and supplier conversations instead.
Whatever fits your stack. We have shipped on PyTorch and open models, on Snowflake and Databricks, on cloud infrastructure inside the client’s own tenancy. The client portal shows the toolkit we vet.
A typical first category is live in eight to twelve weeks from the end of the diagnostic. Rollout to the rest of the catalogue follows once your planners trust it.
You get a ranked list of where AI pays back fastest in your operation, and a build plan for the top item.
Begin free consultation →