Cloud Sustainability Service

The problem

Customers needed to report their carbon emissions, especially with regulation tightening in European markets.
They also needed consistent numbers to see whether their efforts to reduce emissions were actually working.
Figures from the cloud providers were opaque and only available per service or account, and each provider used its own methodology, so numbers were not comparable across clouds.

Actual power data stays with the cloud providers, so the product had to estimate from billing and usage data alone.
Working bottom-up from that data turned out to be an advantage, since it allowed estimates down to each instance, each hour.

A two-step estimate. In-house ML models estimate power use, and emissions (CO2e) are then calculated from that estimate using data center efficiency (PUE) and grid emission factors.
One model per service. Different services have different power characteristics, so each gets its own model.
Explainable and stable over marginally more accurate. Customers report these numbers, so model choice favored explainability and consistent month-over-month results over chasing accuracy with a high-variance model.
Vendor-agnostic methodology. One approach applies across all four cloud providers, which is what makes the numbers comparable.
Built inside Cloudability. Emissions appear next to the cost reports customers already use, with no separate infrastructure to run.

Two data scientists from my team, plus engineering teams for the delivery pipeline.
I conceived the product and built the first version, guided the team in extending the methodology across services and providers, and coordinated with engineering to deliver results end to end.

Bring the other stakeholders in earlier, so the delivery pipeline could have been built in parallel with the methodology and the product reached market sooner.