Status: [UNTESTED] Endpoints found: http://localhost:8000/api/v1/briefs http://localhost:8000/api/v1/briefs/$BRIEF_ID/approve http://localhost:8000/api/v1/briefs/$BRIEF_ID/publish http://localhost:8000/api/v1/cms?brief_id=$BRIEF_ID&status=published http://localhost:8000/api/v1/cms/cms_replace_me/preview Source: https://github.com/farheenfathimaa/brand-content-agent-pipeline # Brand Content Agent Pipeline An end-to-end content pipeline for a **fictional** brand: brief → grounded draft → guardrails → critic revision → human approval → translation with back-translation checks → staged → local CMS publish. Every state transition is persisted; when a webhook URL is configured, each transition is also emitted as an HMAC-signed webhook. The default mode is deterministic **mock mode with zero API keys**. You can run the full workflow, tests, and eval harness without signing up for anything. Live mode adds Groq (primary), Gemini (automatic fallback), and optional Anthropic through LiteLLM. > **Alder Grove is not a real company.** Every brand, product, fact sheet, and brief in this repository is invented for demonstration. [![CI](https://github.com/farheenfathimaa/brand-content-agent-pipeline/actions/workflows/ci.yml/badge.svg)](https://github.com/farheenfathimaa/brand-content-agent-pipeline/actions/workflows/ci.yml) ## The problem Marketing content operations break in predictable places: - A writer or model adds a health claim, spec, certification, or competitor product that the source material does not support. - Brand voice rules live in a PDF, so they are enforced by review comments instead of automation. - Compliance review and publishing happen in separate tools, leaving no reliable audit trail. - Translation quality is assumed until a localized page goes live with a missing disclaimer. - Evals are hand-waved, so nobody knows whether the latest prompt made things better or worse. This project demonstrates a small, inspectable system that makes those failure modes explicit: deterministic guardrails, a bounded revision loop, a human approval gate, a second-language sanity check, and an eval harness that reports real numbers from the code in this repository. ## Architecture ```mermaid flowchart LR Client[Client / curl / UI] --> API[FastAPI API] API --> Pipeline[Pipeline service] Pipeline --> SM[(SQLite\nsta