Projects · 02

aws-platform-terraform

A production-style AWS platform in Terraform, built as reusable modules with tests, one roadmap item at a time.

Read the code

Right now

Checking…

Read from /api/status after the page loads. How it works

My part
It is how I am learning AWS platform engineering. A scheduled Claude Code run builds the next roadmap item and writes a note explaining it; I review each change and study the note.
Result
Every module is tested offline, with no AWS account. GitHub Actions then applies the live stacks to a real account through OIDC, with no stored keys, and destroys them again. More than 20 drill reports with redacted console screenshots are in the repo.
Stack
  • Terraform
  • terraform test with mock providers
  • Moto
  • tflint
  • checkov
  • GitHub Actions with OIDC
A drill report on GitHub titled "Live drill m1, 2026-11-18". Its table reads plan gate passed, apply success in 61 seconds with 39 resources, destroy success in 250 seconds, destroy verified yes.
The report the live drill of 18 November 2026 committed to the repository, as GitHub shows it. Taken with headless Chrome at 1440 by 900, at twice the pixel density.

What it is

Networking, identity, containers, serverless, observability and CI/CD on AWS, written as reusable Terraform modules with tests. Done so far: remote state, VPC, VPC endpoints, flow logs, KMS, GitHub OIDC and IAM permissions boundaries. EKS, serverless and observability are on the roadmap.

How it is built

I am building it with Claude Code as a structured way to learn AWS platform engineering. A scheduled run takes the next item from the roadmap, builds it with tests and pushes it once every check passes. It also writes a learning note: what was built, why, the trade-offs, an exercise and interview-style questions. I review what it built and study the note. Commits written by Claude say so in a trailer.

How it is tested

Every module has terraform test files of two kinds: unit tests with a mocked AWS provider and no network, and integration tests that really create and destroy resources against Moto, a local AWS emulator. CI also runs terraform fmt, validate, tflint and checkov. None of that touches a real AWS account. The unit tests also pass under OpenTofu.

Live drills

Tests prove the code; a live drill proves it works in a real account. After CI passes on a push to main, a workflow applies every stack whose files changed since the last successful drill. It assumes its AWS role through GitHub OIDC, so no AWS keys are stored anywhere. It saves read-only evidence and console screenshots with the account ID redacted, destroys the stack, checks that nothing is left, and commits a report.

The drill runs inside limits it cannot loosen: a plan gate refuses forbidden resource types, the role is boxed in by IAM denies and a permissions boundary, a budget action cuts it off at the spend limit, and a separate reaper workflow destroys anything a drill leaves behind.

The row for this project on the home page is the latest of those drills, read from GitHub’s public API.

Architecture · scroll sideways for all of it →

How aws-platform-terraform is built and drilled A scheduled Claude Code run builds the next roadmap item and pushes to main. CI runs offline: format, validate, tflint, checkov and terraform test with mock providers and Moto. If it passes, a live drill in GitHub Actions signs in to a real AWS account through OIDC with no stored keys. In the account it applies the stack, records evidence, destroys it and checks nothing is left. A drill report is committed to docs/live. Rodrigo reviews the work and the learning note. scheduled run Claude Code builds the next roadmap item push to main CI, offline fmt · validate · tflint checkov · terraform test live drill GitHub Actions AWS account OIDC sign-in, no stored keys in the account apply · record evidence · destroy check that nothing is left drill report committed to docs/live/ with redacted console screenshots review I read what it built and the learning note it wrote
A push to main runs the offline checks. If they pass, the drill applies the changed stacks to a real AWS account, records what it made, destroys it, and commits a report. Drawn from .github/workflows/live-drill.yml in the repository.

Where these claims come from