The system that's supposed to tell your planners what to do —
and doesn't.
Your ERP records what already happened. Your planning tool shows what's happening now. Between them is a human deciding what should happen next, six hundred times a day. Krateos is the autonomous operational layer that runs in that gap, inside your ERP, and makes the decisions the software has never been able to make on its own.
Every manufacturer in the world has the same supply-chain problem.
Your ERP records every transaction in your business. Every purchase, every shipment, every invoice. It tells you what already happened.
Your planning tool, sitting next to it, is a dashboard. It tells you what's happening now.
Between them is a human. Your senior planner. Looking at both, looking at the market, looking at the customers, deciding what should happen next. Six hundred times a day. Every day.
When that planner is good, your business is good. When that planner is tired, your business is tired. When that planner retires, the company pays for it for years.
None of this is a problem your team can fix by working harder. They are already working as hard as human beings can work.
That gap is where Krateos lives.
Two analogies. One product.
We built a team of autonomous agents that run inside your ERP and do the work the software has never been able to run. Today there are eight. Two of them carry the load.
He watches your inventory. When the level drops, he reorders. When it's fine, he doesn't.
His math doesn't need a forecast to work — he runs on reality. Net Flow Position, FEFO-aware on-hand accounting, red/yellow/green zones computed live from MOQ, variability, and lead-time.
She looks at patterns and predicts what's coming. Cold snap tomorrow. Below freezing by 6 a.m.
She doesn't replace Walt — she informs him. 19 forecast models per SKU per cycle. Picks the MASE winner. Ships the reasoning stapled.
You don't replace your thermostat with the weather app. But knowing the weather, you might pre-heat the house tonight.
That's what Dana does for Walt. "I see a 50% demand increase coming on RESIN-4200 in June. High confidence. Pre-heat." Walt temporarily raises the inventory target — say, from 24 barrels to 36 — starting in mid-May. When the June surge arrives, the larger buffer absorbs it. In September, when the season ends, the target returns to baseline.
Every vendor in your market has both pieces. None of them run them in conversation.
They export the forecast to a CSV. They import the CSV into the DDMRP configuration. They run the cycle once a month. They hope nothing has changed since.
- 01Forecasting tool generates a forecast. Once a month.
- 02Planner exports a CSV.
- 03Planner imports the CSV into the DDMRP config.
- 04DDMRP recomputes buffers. Hopes nothing has changed.
- 05Reality moves. Forecast is stale within days.
- 06Buffer reacts to actual demand. Forecast doesn't know.
- 07Repeat next month. Which number do I trust?
- 01Dana refits forecasts nightly. Per SKU. Best-of-19.
- 02Forecast change pushed to Walt. Within seconds.
- 03Walt applies the Demand Adjustment Factor. Pre-heats or cools.
- 04Buffer reacts to actual demand in real time.
- 05Walt's reaction observed by Dana. Models recalibrate.
- 06Both write to the same audit trail. One source of truth.
- 07No CSV. No monthly cycle. One number to trust.
That continuous conversation is the product.
Eight agents. One audit trail. A single shift.
Dana and Walt carry the load. Six others — Brian, Quinn, Sarah, Max, Frank, Grace — pick up the work that follows. By 7 AM there are six decisions in your inbox. Not six hundred.
Two leads. Six specialists.
Dana and Walt run the operational core — forecast and replenishment, in continuous conversation. The rest pick up the work that follows. Each named, role-scoped, and authority-limited. Trust widens only as they earn it.
The forecasting depth no one else ships.
19 models benchmarked per SKU per planning cycle. SeasonalNaive is the floor — no forecast ships unless it beats naive on hold-out MASE.
Eight models. One winner.
- autoAutoARIMA · AutoETS · AutoTheta · DynamicOptimizedTheta
- seasonalMSTL · Holt-Winters · SeasonalExponentialSmoothing · SeasonalNaive
- selectionHold-out MASE bake-off. Refit on full history.
Low-velocity SKUs don't get faked.
- familyCrostonSBA · CrostonClassic · ADIDA · IMAPA · TSB
- routed5 archetype classes pick the family. Trending · seasonal · multi-seasonal · intermittent · spiky.
- gateMust beat naive on MASE. Else the forecast doesn't ship.
LightGBM, three paths.
- per-SKULag features [1, 4, 13, 52] + exogenous regressors.
- batchShort-history SKUs borrow strength from similar ones — the primary ML advantage.
- bake-offHead-to-head stats vs. ML. MASE winner picked. No shrugging.
The sum of the parts equals the whole.
- hierarchyCustomer × Region × SKU reconciled via Nixtla's open-source solver.
- shapSHAP feature importances when ML wins. You know why.
- shippedArchetype · MASE · CV window · importances. No black box.
DDMRP, canonical — not "inspired by."
Demand-Driven Material Requirements Planning, implemented to the spec. Red / yellow / green from MOQ, variability, lead-time, and cycle factor. Live today across 31 SKUs.
Net Flow Position, FEFO-aware. What's actually available after allocation and expiration — not what the ERP thinks is in the rack.
- Three profile archetypes · LONG-HIGH-MOQ, MED-MED-MOQ, SHORT-LOW-MOQ
- Demand Adjustment Factor — symmetric. Saves by not buying.
- γ-reprofile — canonical profile-clone-and-reassign. No parallel override files.
- Shelf-life ceiling cap — no buffer bigger than what will consume before expiration.
- Aging-in-red — not "in red" but "in red for N days." Severity operators can prioritize.
- Confidence-gated auto-propose — bad signal never becomes an auto-action.
- Atomic writes. Overlap rejection. Baseline frozen at approval. Audit integrity by construction.