carbon accounting for ai inference

Built a platform that measures the carbon emissions of AI inference and offsets them as part of the bill. Every model call has an energy and emissions cost. Most stacks never measure it. This one meters it, prices it, and produces a receipt you can audit.

what it does

The platform sits alongside an inference workload and tracks emissions per request through a telemetry pipeline. It manages the offset purchases that retire those emissions, then handles billing and receipt generation so the carbon cost lands on the invoice instead of in a sustainability report nobody reads.

Stripe webhooks process the payments. The system generates PDF receipts with digital signing, so a customer ends up with a tamper-evident record of what they emitted and what was retired against it.

the data model

Underneath is a relational model of eleven entities covering workloads, telemetry, offsets, billing, and receipts. The design treats traceability as the core requirement. Every emission number connects back to the request that produced it, and every offset connects back to the emissions it cancels. That chain is what makes the receipt defensible.

how it is built

The architecture runs on AWS with a documented REST API and a clean separation between the telemetry pipeline, the billing path, and the receipt service. The documentation is built as a single source of truth: product proposals and user stories, domain specs and API contracts, a system design with a numbered decision log, and the implementation roadmap. Diagrams carry the system context and the data flows.

why it matters

AI compute is becoming a real share of data center load, and the emissions attached to it are mostly invisible at the point of use. A team that wants to account for that today has to bolt together cloud bills, grid intensity data, and a spreadsheet. This collapses the loop. Measure the inference, price the carbon, retire the offset, sign the receipt. The emissions become a line item with an audit trail instead of an estimate in a slide.