✓

Describe

Review generated

Run baseline

Review what we'll build

From your description, ownEvo generated a simulator, an eval set, a success metric, and a UI layout. Each is editable. Once you approve, the agent runs against the simulator to establish a baseline — then the improvement loop starts.

From your
description

"Forecast weekly demand at SKU-store level for our 8,400 SKU catalog across 142 stores. Flag SKUs likely to need markdown within 4 weeks. Account for seasonality, supplier price-build, and regional variance. Pull historical sales from SAP ERP. Pull current weather forecasts from NOAA…"

Edit ↗

Simulator

6 tools · 2 user behaviours · 3 environment generators

Tools the agent can call

forecast_demand(sku, region, weeks_ahead)

Returns a forecast with confidence intervals.

derived from "Forecast weekly demand at SKU-store level"

Edit

fire_markdown_alert(sku, region, severity, recommended_discount_bp)

Notifies the supply chain VP. Logged for audit.

derived from "Flag SKUs likely to need markdown within 4 weeks"

Edit

query_sap_sales(sku, region, start_date, end_date)

Historical sales by week. Live in production; replayed deterministically in simulator.

derived from "Pull historical sales from SAP ERP"

Edit

fetch_noaa_weather(region, weeks_ahead)

Forecast deltas from NOAA, normalized to anomaly score.

derived from "Pull current weather forecasts from NOAA"

Edit

list_skus(category?, region?) · get_inventory_status(sku, store)

Catalog and inventory lookups.

Edit

User behaviour (simulated)

Supply chain analyst · Monday morning markdown review

Opens markdown report 9:00 PT each Monday, asks the agent to forecast next 4 weeks per region, drills into anything flagged high severity.

inferred from supply chain forecasting domain pattern

Edit

Supply Chain VP · daily alert triage

Reviews fired markdown alerts daily, approves/rejects each, comments form new eval cases.

derived from "The supply chain VP reviews markdown alerts daily"

Edit

Environment generators

Synthetic SKU catalog · 8,400 SKUs across 6 categories

Mirrors your real catalog distribution: footwear (winter/summer), apparel, accessories, etc. Names and IDs randomized.

Supplier behaviour · 12 suppliers, 3-12wk lead times, monthly price drift

Reproduces the supplier price-build dynamic mentioned in your description.

derived from "Account for seasonality, supplier price-build"

Weather generator · NOAA-shaped anomalies

Generates realistic weather forecasts with anomaly events (PNW cold snaps, Southeast hurricanes, etc.) for cold-weather category testing.

derived from "Pull current weather forecasts from NOAA for cold-weather categories"

Eval cases · 12 generated

3 from your past misses · 4 regression checks · 5 generated edge cases

Case

Type

Severity

PNW winter boot 2025-W47 (replay)

From: "missed the 2025 Pacific NW winter boot spike by 4 weeks"

Past miss

High

Edit

Promo bundle uplift Q4 2024 (replay)

From: "underweight promotional uplift on bundled SKUs"

Past miss

Medium

Edit

Promo bundle uplift Q4 2025 (replay)

From: "underweight promotional uplift on bundled SKUs"

Past miss

Medium

Edit

Steady-demand baseline (no anomaly)

Regression: agent must NOT fire false alerts when demand is normal

Regression

Low

Edit

Inventory-pegged forecast (low stock)

Regression: forecast should not exceed available inventory under any signal

Regression

Low

Edit

Cold snap surge · winter coats Northeast 2024-W44

Edge case: weather-driven anomaly, narrow window

Edge

Medium

Edit

Supplier delay · 12wk lead time SKUs

Edge case: supply-side disruption affects forecast accuracy

Edge

Medium

Edit

Regional variance · Southeast hurricane evacuation

Edge case: localized demand displacement

Edge

High

Edit

9-12

+ 4 more cases (new SKU intro, EOL, returns spike, holiday shift)

Standard supply-chain edge cases for the catalog

Mixed

—

View

Success metric

Composite of precision, recall, and discount accuracy

metric: markdown_window_composite precision = correct_alerts / total_alerts_fired recall = correct_alerts / total_actual_markdowns discount_bp_accuracy = mean( |recommended_discount − actual_discount| <= 5pp ) correct_alert := alert fires within 6 weeks of actual markdown event AND alert.severity matches actual markdown size band composite = 0.5 × precision + 0.3 × recall + 0.2 × discount_bp_accuracy # derived from your description: "an alert is correct if it lands # within 6 weeks of the actual markdown event and the recommended # discount is within 5pp of what we ultimately ran"

Higher is better. Range: 0.0 – 1.0 · Computed per eval case · aggregated as mean across 12 cases · Reported as % in dashboards

Operate-view UI · 4 primitives selected

What the analyst will see when operating this workflow

Preview layout →

MetricCards

TimeSeriesChart

TableView

AlertList

KanbanBoard

ConversationView

‹ Back to description

Run baseline →