Numbers Finance Can Trust
Sizing from p95 usage with variability-adaptive headroom, a 70% recovery-rate assumption, tiny findings suppressed, and unmeasured metrics excluded rather than zero-filled: no fabricated waste on machines nobody monitored.
CASE STUDY
GCP Spotlight
Waste, found and finished.
An organization-wide GCP optimization platform covering Kubernetes, virtual machines, storage, and AI spend. It sweeps the fleet twice a day, prices every finding from real billing data, and then does what most optimization tools never do: tracks each recommendation to closure, and only claims a saving when the bill actually drops.
$117.5K
per month of itemized savings identified
1,957
recommendations against named resources
$2.0M
per month of GCP spend monitored
70%
recovery rate assumed, never 100%
01
The problem
Optimization tooling has a credibility problem. Dashboards shout theoretical savings computed at best-case rates, count the same waste twice, invent utilization for machines nobody ever measured, and never follow up on whether anything was actually fixed. Engineering teams learn to ignore them in about a month. The design goal here was the opposite: a number conservative enough to put in front of finance, and a system that closes its own loop.
02
The system
Twice a day, Spotlight collects the entire fleet: Kubernetes workloads, VM utilization, storage buckets, AI model spend, into isolated per-service stores. Pricing is billing-first: figures derived from the actual bill are labeled as such, and anything estimated says so on its face. Access is read-only by design: workload identity with no service-account keys anywhere, and a custom storage role that can read bucket metadata but is incapable of reading a single object's contents. And findings ship as artifacts, not advice: exportable Kubernetes patch YAML and storage lifecycle JSON, ready to apply.
03
The recommendation lifecycle
CLOSING THE LOOP
Savings Are Claimed When the Bill Drops, Not When the Slide Ships
Every recommendation gets a stable fingerprint and a state machine: new, accepted, implemented, dismissed, or expired. "Implemented" is never self-declared: it's auto-credited only when the resource's observed cost actually falls to within 10% of the target, banked against a frozen baseline. A recommendation that quietly disappears without ever being accepted expires with zero credit. The result is two honest, strictly separate numbers: what could be saved, and what has been.
CONSERVATIVE MATH
SECOND OPINION
Google's Recommenders, Cross-Checked
Seven of Google's own Compute recommenders run alongside Spotlight's sizer as an independent check. Agreement raises confidence, disagreement earns a look, and overlapping findings are never double-counted.
DATA INTEGRITY
Failure Can't Delete the Truth
Every collection records success or a classified failure reason, and cache pruning only touches scopes that read successfully: an API outage can age the data, but it can never silently erase good rows.
04
The stack
GCPGKEGCEGCS
Cloud BillingCloud MonitoringFastAPIPython
ReactTypeScriptTailwindSQLite
Workload Identity
05