The number on top is money found, not money spent.
Every screen below is the actual product rendering a synthetic ~$85k/month estate (illustrative — no customer data). The engine attributes every dollar through lineage, then turns what it finds into priced actions, receipts, and one-page reports.
Daily estate cost — one bad day, named
Anomalies are median/MAD outliers per data product — and every flag names the run that caused it, because attribution is lineage-derived, not workspace-level.
Recommended actions
| action | subject | window | projected | confidence | effort | |
|---|---|---|---|---|---|---|
| unallocated_closure $10,289.57 unexplained over 30d — closure checklist | platform | $10,289.57 | $125,190/yr | estimated | M | copy fix |
| idle_warehouse idle $4,500.82 of $9,582.02 warehouse spend (47.0%) | wh-bi-02 | $4,500.82 | $54,760/yr | strong | S | copy fix |
| failed_run_waste $1,811.68 spent on runs that produced nothing | j-1009 | $1,811.68 | $22,042/yr | strong | M | copy fix |
| duration_regression runs above 1.5× trailing median since 2026-07-24 | j-1013 | $47.69 | $580/yr | estimated | M | copy fix |
| low_utilization avg cpu 12% — downsize candidate (classic compute) | 0721-093015… | $81.86 | $996/yr | estimated | M | copy fix |
Six recommendation types: idle warehouses, failed-run waste, duration regressions, low utilization, spin-up overhead, unallocated closure. Every row carries evidence, projected $/yr, confidence, effort, and a copyable remediation.
Data product P&L — cost × observed consumption
| data product | projected | readers 90d | observation |
|---|---|---|---|
| lake.customer_360 | $115,047/yr | 9 | last read observed 2d ago |
| lake.events | $97,861/yr | 24 | last read observed 3d ago |
| lake.observability | $46,566/yr | 0 | no reads observed in 90 days |
| lake.exec_kpis | $44,162/yr | 8 | last read observed 5d ago |
The sensor is honest: coverage shown (here 92.0% of attributed cost has ≥1 observed read), unobservable read paths listed, and the product never says a table is safe to drop — it says "no reads observed" and hands you verify-before-drop steps.
Receipts — measured, not promised
found $258,899/yr · realized (measured) $30,077/yr
Every recommendation has a lifecycle — detected → acknowledged → actioned → realized. Realized dollars come only from measured before/after rates across engine runs. Projections never book. That's your renewal case, written by the engine.
Roadmap: AI-assisted recommendations via Databricks-hosted model serving — analysis stays inside your workspace, same zero-egress posture.