How the engine works.

Every brief starts with your own merchant history — units sold, by SKU, by week, by category. Not a peer aggregate, not a benchmark, not what a global retailer was buying in Antwerp last autumn. We model each garment's demand from the data your own floor has generated, on the customers your own brand actually serves.

Where the data is thin — a new line, a new SKU, a new shoulder window — the brief says so out loud. The engine never extrapolates further than the signal allows, and it tells the buyer when the band is wide. No confident guess passing as certainty.

The Indian commercial calendar is a forty-two-window forcing function on demand. Diwali bends colour. Karva Chauth bends silhouette. Navratri bends print. The wedding season bends everything. A forecast that does not know the calendar is a forecast that does not know the country.

Each window in our calendar carries a category multiplier — what lifts, by how much, in which region — measured on real Indian D2C floors and re-checked every season. Those multipliers feed directly into the demand model. The buyer sees the multiplier on the page, not hidden in the maths.

Cultural signals matter, but the signal must be intentional. The engine reads runway looks from a curated list of designers who actually influence the Indian D2C buyer. It reads competitor catalogues on the Indian D2C floors that share your customer. It reads Indian media coverage of the season ahead. Every signal is named, scoped, and weighted.

We use embedding-based similarity to attach each runway look and each competitor SKU to an attribute in our taxonomy. The result is a heat map of what the market is leaning into, expressed in the same language the buyer uses for the order book. No black box, no “the data says so” without a path to which data and where it came from.

Forecasts have prediction bands for a reason. We use an ensemble of forecasting techniques, and we use the disagreement between them to compute an honest range — a calibrated prediction interval — rather than a single number with false precision. When the techniques agree, the band is tight. When they disagree, the band widens, and the brief tells you why.

Every account's actual accuracy is reported on the page. Forecast miss for the last quarter, broken down by category and window. The engine's own confidence is held to the same standard it asks of the buyer's instinct: show your work.

The output is a buying brief in a senior buyer's voice — Anaita, the composite character we introduced in the manifesto. Every line in the brief is grounded in a number on the page: a sell-through, a prediction interval, a runway analog, a competitor sell-out, a festive multiplier. She names the call, names the risk, and shows the maths underneath.

When the signal is thin, the brief says watch, don't load and names what would change its mind. When the signal is strong, it gives a depth-of-buy range that accounts for unit margin, markdown risk and salvage value — not a single confident number. The engine is an opinion you can interrogate, not an oracle you have to trust.