◆ CHIHUAHUAS STUDIOS
PrivacidadeDados locais
◆ CHIHUAHUAS STUDIOS

CHIHUAHUAS STUDIOS • ENGINEERING PRODUCT

DEVOPS
AI AGENT

A safe, evidence-driven DevSecOps validation agent. Deterministic checks own the decision; AI interprets bounded evidence but cannot turn a deterministic failure into PASS.

CHIHUAHUAS STUDIOSENGINEERINGV0.1.0PRIVATE REPO • PORTFOLIO PREVIEW
200TESTS IN VERIFIED E2E PROOF11 application + 189 Agent at validated run
0VULNERABILITIESreference validation
OIDCAWS AUTHENTICATIONno static access keys
PASSFINAL DECISIONdeterministic authority

THE IDEA

“It is not a pipeline that calls ChatGPT. It is a deterministic DevSecOps pipeline with a controlled, auditable AI analysis layer that cannot reduce a deterministic failure.”

TECHNOLOGY

GitHubRepos • Actions • Artifacts
AWSOIDC • STS • CFN • S3 • CloudFront
OpenAIResponses API • structured analysis
PyPython 3.12Typer • Pydantic • pytest

END-TO-END ARCHITECTURE

Trust boundaries are part of the product.

The untrusted pull-request zone has no production secrets. Deterministic evidence crosses into a trusted default-branch workflow. Only there can bounded evidence reach OpenAI. AWS observation and deployment use separate roles.

01 / UNTRUSTED → DETERMINISTIC

VALIDATION PIPELINE

Pull requests never receive production credentials or authority to change the final result.

INPUT

GitHub repository

Next.js reference application, pull request and workflow trigger

QUALITY GATE

Deterministic checks

Tests, dependency audit, build and bounded structured evidence

ARTIFACT

Evidence only

Sanitized output; untrusted claims cannot override the decision

SECURITY BOUNDARYNo OpenAI secret or AWS deploy credentials in pull-request execution.
02 / TRUSTED DEFAULT BRANCH

CONTROLLED AI ANALYSIS

Trusted workflow reads bounded evidence and optionally requests structured interpretation.

TRUSTED TRIGGER

workflow_run

Trusted Agent checkout; evidence revalidated before interpretation

OPTIONAL EXTERNAL API

OpenAI Responses

Bounded evidence in; findings and risk assessment out

FINAL AUTHORITY

PASS / WARN / BLOCK

Deterministic checks own the outcome; JSON, Markdown and HTML reports

TRUST BOUNDARYAI can explain risk, but cannot turn a deterministic failure into PASS.
03 / SEPARATE AWS RESPONSIBILITIES

CLOUD & DELIVERY

Site publishing and AWS observation are independent permission paths.

DEPLOYMENT ROLE

GitHub Actions → OIDC / STS

Temporary deploy credentials

S3 private bucket → CloudFront + OAC

Static export, cache invalidation and HTTPS delivery

READ-ONLY ROLE

Agent → OIDC / STS

Separate temporary credentials

AWS inventory → bounded evidence

Explicitly allowed observations of CloudFormation, S3, CloudFront, Route 53, ACM and IAM

INFRASTRUCTURE AS CODECloudFormation templates describe intended infrastructure; live configuration must be reconciled before applying changes.
Deterministic evidence Trusted AI interpretation AWS temporary credentials
01

Deterministic first

Reviewed adapters execute bounded checks for Next.js, Terraform and CloudFormation. Mandatory failures remain failures regardless of AI output.

02

Trusted AI boundary

OpenAI receives sanitized, bounded evidence from a trusted workflow. It returns structured risk, findings and recommendations — never commands to execute.

03

AWS without static keys

GitHub Actions OIDC obtains temporary STS credentials. The Agent collector is read-only and separate from the website deployment role.

04

Auditable evidence

Each run can produce JSON, Markdown and HTML with checks, test inventory, dependency audit, AI audit trail and final PASS / WARN / BLOCK.

SOURCE CODE

Built as a real system.
Presented as engineering proof.

REPOSITORY (PRIVATE) ↗