Demand & Assortment Intelligence

One cadenced pull, never a stampede

Connectors own every external fetch on a schedule. The workspace reads from stores — it never hammers a source system, and a source that goes dark degrades the answer honestly instead of silently faking it.

• Sales & transactionsUnit and revenue movement by SKU, store, and channel — the ground truth everything else is measured against.
• Inventory positionsOn-hand, in-transit, and on-order by location, with the as-of timestamp preserved so staleness is visible.
• Product catalogHierarchy and attributes — category, style, size, color, price band, season.
• Supplier commitmentsLead times and open purchase orders, which set what a reorder can actually promise.
• Engagement signalsSearch, page, and cart activity — demand intent that shows up before it shows up in sales.

Structure the Brain can reason across

Retail questions are multi-hop by nature: this style is dying — what substitutes it, which suppliers are exposed, which stores hold the depth? That's a graph traversal, not a table scan. Every node carries its source and as-of time, so an answer can cite where each number came from.

• EntitiesSKU, Style, Category, Store, Channel, Supplier, DemandSignal, InventoryPosition, PricePoint, Promotion.
• RelationshipsSKU→Style→Category · Supplier supplies SKU · SKU substitutes SKU · Store→Region · Promotion covers SKU.
• ProvenanceEvery fact is tagged live / store / cached / simulated. A missing datum is excluded from the answer, never imputed to fill a gap.
• Semantic layerEmbeddings over products and merchant notes, so a plain-language question retrieves the right slice before reasoning starts.

Components, then a cross-sectional rank

Independent components each score one thing they can defend. They're combined and then ranked cross-sectionally within category and horizon — not scored on an absolute scale.

That detail is a scar, not a preference. The financial-services workspace first shipped absolute scores and every qualifying name pinned to the cap, flattening exactly the differentiation the score existed to produce. Relative ranking is what makes the output actionable.

• Sell-through velocityRate versus plan and versus comparable items at the same lifecycle week.
• Stockout exposureProjected lost units where demand is running ahead of cover.
• Curve healthSize and color integrity — a style with a broken size curve is already dead at full price, however healthy the aggregate looks.
• Markdown riskAged units and terminal-inventory exposure at the current sell-through rate.
• Elasticity & promo liftMeasured response to past price and promotion moves, not an assumed curve.
• Launch rampNew-product trajectory benchmarked against analogous historical launches.
• Placement imbalanceSame style selling out in one region while sitting in another — the transfer opportunity.

Every call carries a horizon and an outcome

A signal that is never scored is marketing. Each call is written to an outcome ledger with its barriers fixed at the moment of the call, then resolved on what actually happened. Hit rates accrue only from closed rows — so a track record can't be inflated by counting the open ones.

• BarriersReached target sell-through · forced into markdown · went to stockout — whichever comes first inside the horizon.
• Closed-only ratesOpen calls are excluded from every published rate. No projected or in-flight outcomes.
• Learned weightsComponent weights update from closed outcomes within bounded limits, so the model adapts without drifting.
• Versioned labelsChange the grading rule and it becomes a new label version; rates re-accrue rather than silently mixing methodologies.