Ask for a dashboard, watch it build: analytics for 400+ warehouses.

A network of more than 400 warehouses produces a lot of operational data: daily inventory snapshots by commodity, inward and outward transactions, daily capacity utilisation and monthly revenue split into storage, handling and insurance charges. The workspace was built to show what it looks like when operations and finance staff can question that data directly. Instead of filing a request for a new report, a manager types what they want to see and gets a working dashboard they can change, save and come back to.

The screen is split in two: chat on the left, a live canvas on the right. A request such as a six-month revenue projection goes to a Claude agent running a tool-use loop with nine tools, and panels stream onto the canvas as each query or forecast completes. Follow-ups like "add a panel showing the top five warehouses by revenue" or "change panel 2 to a line chart" patch the dashboard in place, and any dashboard can be saved as a report, reopened and updated.

Forecasts come from a separate Python service using Holt-Winters exponential smoothing, and each forecast panel shows 95% confidence intervals so planners can see the range, not just a single line. The same service exposes endpoints for anomaly detection with Isolation Forest, seasonal decomposition and correlation between series. If the service is unavailable, the workspace falls back to a linear regression forecast rather than leaving the panel empty.

The agent does not write free-form SQL for the core figures. A semantic layer defines 11 governed metrics, 12 dimensions and 10 time-range presets, and compiles each request into deterministic SQL, so two people asking the same question get the same number. A metrics page shows the full catalogue, which gives finance and operations one place to agree what each figure means. Server-side output guardrails clean every agent response before it reaches the screen.

The workspace was built as a demo and runs on seeded data in a warehousing schema modelled on the operator's network, not on its production systems. It connects to PostgreSQL and includes a schema browser, so pointing it at a real database is a connection change rather than a rebuild. No usage or outcome figures exist yet, because it has not been deployed to the operator's teams.