The autonomous AI demand planning platform

Yes. Demand forecasting is the core of the platform: forecasts by SKU and location, reconciled hierarchically and scored against your own history. The S&OP and IBP cycle and inventory optimization run on that same number. It is a SaaS product built in Argentina, in Spanish and English.

No. There is no implementation fee and no professional-services line. The pilot is the implementation: we load your history, wire the connectors and hand you a plan, and if you stay it carries straight into the subscription.

Sales history by SKU and location is enough to start forecasting — two years is the ideal, twelve months will do. Current stock is what lets us repair the weeks the shelf was empty, so a stockout stops teaching the model that nobody wanted the line. The modules downstream need their own inputs to be modelled properly: lead times and costs for replenishment, capacities and changeovers for scheduling, addresses and time windows for routing. A CSV export works; a connection to your ERP works better.

ERPs, ecommerce and spreadsheets: SAP, Dynamics, Odoo, Tango, Mercado Libre, Tiendanube, Shopify, Google Sheets. If yours is not on the list, the connector is something our agents write, which is the point of having them.

Gradient boosting (LightGBM, XGBoost), the classical statistical family (ETS, ARIMA, Theta, seasonal naive as the reference everything is scored against), neural networks built for series (N-BEATS, NHITS, TFT, LSTM), pre-trained foundation models for time series (TimeGPT, Chronos, TimesFM, Moirai) that give a usable forecast on a line that barely has a history yet, intermittent-demand models (Croston, TSB) for anything that sells in bursts with long gaps between them, and global models trained across your whole catalogue for the lines with too little history of their own. New ones go on the bench as they are published — and every week we fit more than a thousand variants per customer, so each series ends up with whichever one measures best on it, not whichever one was fashionable when you signed.

Yours first: sales by SKU-location, price and discount, promotions, the stockouts we repaired, lead times, your commercial calendar, and the product and channel hierarchy. The models also read the shape of the series itself — the trend, the seasonality of the year and the one inside the week, and the level a line settles at after something changes. Then the outside world: holidays, weather, FX, search interest, competitor prices, sector indices. A variable only stays in if it lowers the measured error on your own history, which is why most of them do not.

Your team, every time. The agents propose and explain; a person approves. Every override is stored with its reason and its author, and scored against what the model would have done — which is how the platform learns who to listen to.

It gets measured. Every run is scored against what actually happened, on the same reference, and you see the error per line before you commit to anything. A series we cannot grade is labelled ungradable rather than given a number.

No. The models, the features and the constraints are the agents’ work, reviewed by our team. What your planner brings is the judgement about your business, which is the part no model has.

Never for anyone else. Your series stay in your project, the models are fitted on your data for you, and nothing from your operation improves anything of ours or of another customer’s.

Full export in open formats at any time, deletion within thirty days of asking, backups included, with written confirmation. Monthly plans cancel monthly.