missionfinops/local-ai-finops

This repository is not a piece of software you install and run. It is a written architecture proposal from Mission FinOps describing a possible design for using AI inside large companies to investigate cloud spending, while keeping sensitive company data under the company's own control rather than sending it to an outside AI service. FinOps refers to the practice of managing and explaining cloud computing costs. The document itself says clearly that this is a concept, version 0.1, and that not every piece described in it currently exists. The core problem the document lays out is that explaining a change in cloud costs usually requires pulling together many different kinds of internal information: billing data, business context, engineering changes, support tickets, ownership records, accounting details, and internal policies. For a regulated company, sending all of that sensitive information to an external AI platform may not be allowed. The proposal argues the real challenge is not picking which AI model to use, but building a trustworthy system around it that can show its evidence. The proposed design runs the AI model close to the company's own data, inside the company's own boundary. The AI model itself is treated as replaceable and can be swapped for any approved model, open source or otherwise. What stays constant are four building blocks: tools that collect and organize evidence, skills that represent repeatable investigation steps, a layer of business specific context, and a layer of general FinOps knowledge. The model's job is to interpret and explain the evidence, not to be treated as the source of truth itself. The document points to one existing open source project, called Kulshan, as an early building block for making AWS cost investigation more local and repeatable, though it notes Kulshan does not implement the full architecture described here. The proposal repeatedly stresses that any explanation the system produces should show its supporting evidence, note contradicting evidence, state its confidence level, and be honest about what remains unknown, rather than claiming certainty it cannot support.