ContentForge is an autonomous, multi-agent AI pipeline designed to scale high-quality SEO content generation while strictly enforcing enterprise compliance and monetization constraints (such as Google Ads policies).
Built entirely with a LangGraph state workflow and powered by Google Gemini models, ContentForge moves beyond single-purpose AI wrappers to showcase true, deterministic agentic orchestration.
The core of ContentForge is a 5-node LangGraph state workflow. Each node represents a distinct AI agent with a specific role.
Below is the deterministic flowchart of the pipeline. Notice how the Decision Gates (Validator and Reviewer) act as automated AI Evals (LLM-as-a-Judge) to route or reject content:
graph TD
A[Raw Keyword Batch] --> B[Scorer Agent]
subgraph Multi-Agent State Workflow
B -->|If Viable| C[Planner Agent]
B -.->|If Rejected| Z[Drop Keyword]
C -->|3 Outlines| D{Validator AI Eval}
D -.->|Fail| Z
D -->|Pass| E[Generator Agent]
subgraph Local RAG
db[(ChromaDB: Policy Rules)] -.-> E
end
E -->|Draft Article| F{Reviewer AI Eval}
end
F -->|Pass| G[CMS / Master Feed]
F -.->|Fail| H[Failure Dashboard]
%% Styling the nodes for visual clarity
style B fill:#3b82f6,stroke:#1e40af,color:#fff
style C fill:#3b82f6,stroke:#1e40af,color:#fff
style D fill:#8b5cf6,stroke:#5b21b6,color:#fff,shape:diamond
style E fill:#3b82f6,stroke:#1e40af,color:#fff
style F fill:#8b5cf6,stroke:#5b21b6,color:#fff,shape:diamond
style db fill:#10b981,stroke:#047857,color:#fff
- 🎯 Scorer Agent: Evaluates raw keywords for commercial viability and search intent. Rejects low-value keywords immediately to save compute.
- 🗺️ Planner Agent: Generates 3 distinct, high-quality SEO content outlines (angles) for approved keywords.
- ⚖️ Validator Agent: Acts as an internal filter, reviewing the generated outlines to ensure they are logical, non-spammy, and structurally sound.
- ✍️ Generator Agent (with RAG): Queries a local ChromaDB vector database containing strict policy documents to ground its generation. It drafts full-length articles tailored to the validated outlines.
- 🛡️ Reviewer Agent: The final compliance check. Grades the output against policy rules. Articles that fail are flagged with exact reasoning and blocked from the final export feed.
- Orchestration: LangGraph (StateGraph, Conditional Edges)
- LLMs: Google Gemini (2.5 Pro for reasoning/validation, 2.5 Flash for high-speed generation)
- API: FastAPI (with Server-Sent Events for real-time streaming)
- RAG / Vector DB: ChromaDB + LangChain Google GenAI Embeddings
- Framework: React + Vite
- Styling: Tailwind CSS v4 + Framer Motion
- Features: Real-time state visualizer, autonomous batch processing, and a high-level execution dashboard.
- Python 3.9+
- Node.js (v18+)
- Google Gemini API Key
Navigate to the backend directory, create a virtual environment, install dependencies, and start the FastAPI server:
cd backend
python -m venv venv
source venv/bin/activate
pip install -r requirements.txt
# Start the SSE streaming API
uvicorn main:app --reload --port 8000Navigate to the frontend directory, install dependencies, and start the Vite dev server:
cd frontend
npm install
# Start the React UI
npm run dev- Open your browser to
http://localhost:5173. - Enter your Gemini API key in the top navigation bar.
- Click "Generate Demo Batch" to simulate a keyword feed.
- Click "Run Entire Batch Automatically" and watch the LangGraph state machine route the data through the agents in real-time.
While this repository serves as a pilot, scaling to thousands of articles requires moving from prototype to enterprise-grade infrastructure:
- Vector Database: Swap local ChromaDB for managed Pinecone or Weaviate.
- Observability: Integrate LangSmith or Arize Phoenix to monitor token usage and trace agent hallucinations.
- Resilience: Wrap the LangGraph execution layer in Temporal.io or AWS SQS to handle async queueing and rate-limiting gracefully.
Built with ❤️ alongside AntiGravity (powered by Gemini 3.1 Pro).