infra-whisperer

Infra Whisperer is an agent that watches a live AWS environment, diagnoses failures by cross-referencing CloudWatch, logs, ECS events, and Terraform state, explains the diagnosis in plain English, and proposes a fix as a pull request. A human must merge the PR and run terraform apply; nothing is changed automatically. It was built as a demonstration of forward-deployed engineer work rather than a chatbot demo. The system uses a deterministic, evidence-based confidence rubric rather than model judgment. The rubric scores hypotheses based on alarm correlation, Terraform confirmation, log matches, ECS events, independent second signals, temporal coincidence, and contradicting evidence. Verdicts range from HIGH to REJECT, and the rubric has been validated against the project's real incident history. It also includes a semantic Terraform state parser that understands inline versus separate security group rules to prevent false drift reports. The project includes production safety infrastructure such as S3 and DynamoDB remote state with locking, scoped IAM roles, and a multi-step approval workflow. It is cost-capped by design with an AWS Budget alarm and one-command teardown. The agent opens a PR with a Terraform diff and a plain-English explanation; a human merges it. Live testing surfaced findings about monitoring gaps, agent blind spots, reasoning failures, confidence calibration, and Terraform state parsing.