A platform sandbox and AI-native tooling for a financial services engineering organisation
Client context
An Australian cryptocurrency exchange operating a private development platform on AWS. The platform team owned a large infrastructure monorepo (10,000+ commits) covering Terraform and Terragrunt IaC, Kubernetes, GitOps, and application configuration for the whole engineering organisation.
The problem
Infrastructure changes were being designed against the main company tenancy, where the blast radius of a mistake is the production platform itself. The team needed a way to prototype significant platform capabilities (new cluster architectures, GitOps topologies, CI runner models) safely, prove them, and then promote them into the main monorepo. Alongside this, service knowledge was scattered across Confluence, Jira, and GitHub, with no unified catalog answering “who owns this and where does it run”.
What was built
- An isolated platform sandbox tenancy: a separate AWS tenancy for testing infrastructure changes, with on-demand ephemeral EKS clusters destroyed nightly. Built with Terraform and Terragrunt (roughly 40 reusable modules), Argo CD per cluster, Karpenter, Bottlerocket nodes, AWS Account Factory for Terraform, SOPS secret management, and GitHub Actions with OIDC-scoped IAM. The sandbox tracked which capabilities graduated into the main monorepo.
- A Backstage service catalog, including a custom plugin, plus a Tilt-driven local platform bringing up Tekton, Argo CD, Backstage, and a Go Kubernetes operator on a laptop.
- An AI-native layer over the platform: a production-shaped MCP (Model Context Protocol) reference architecture in TypeScript; a natural-language interface to the Backstage catalog with strict read-only guardrails; and a knowledge service indexing Confluence, Jira, and GitHub, linked to the catalog and exposed through its own MCP server and an AWS Bedrock chatbot.
- CI groundwork: a self-hosted GitHub runner proof of concept (Actions Runner Controller) and a spike translating GitHub Actions pipelines to Tekton.
Outcomes
- Incidents traced to infrastructure changes fell, and CI/CD became substantially faster through unified configuration and CI reporting native to GitHub, replacing reporting that previously sat detached in a third party CI provider.
- Capabilities promoted from the sandbox into the main tenancy so far: OIDC scoped IAM, so each pipeline receives only the permissions it needs, and a dynamic IaC workflow in which static pipelines automatically discover new infrastructure components in the repository.
- The knowledge service is now part of the delivery loop itself: engineers build features through IDEs and agent harnesses connected to it over MCP, retrieving context from GitHub and Atlassian sources as they work.