CongKong: AI Content Factory

Generating high-volume, search-optimized technical content is incredibly slow and expensive when done manually. However, simple one-shot LLM prompts fail because the AI produces generic, dry text and suffers from factual hallucinations (incorrect tech facts or broken code snippets).

To be valuable in production, AI-generated content needs to be factually accurate, structured, SEO-optimized, and automatically formatted for publishing without constant manual human intervention.

As a Founding Engineer, I built an autonomous multi-agent content factory powered by the Gemini API and Supabase:

• Multi-Agent Orchestration: Created distinct AI agents (Research, Technical Writer, and SEO Expert) that cooperate to process source ideas into structured posts.
• RAG Validation Pipeline: Integrated a Retrieval-Augmented Generation check using vector embeddings. The pipeline compares the output against a curated knowledge base of trusted documents to detect and discard hallucinations.
• Auto-Deployment Automation: Utilized n8n and WordPress REST APIs to parse markdown outputs, format tables and images, inject SEO tags, and publish posts automatically.

Our initial agent pipeline struggled with conversational drift and infinite validation loops. The Technical Writer Agent and the SEO Agent would pass revisions back and forth indefinitely, bloating token costs and degrading article quality.

I resolved this by refactoring the system into a Directed Acyclic Graph (DAG) state machine using LangGraph. I defined rigid input and output schemas (zod structures) for each node.

Under this model, the SEO Agent is limited to a single validation pass, generating structured modification instructions. The final article must pass through a strict semantic evaluator node. If the semantic similarity score between the generated draft and the source documentation falls below 85%, the system halts, logs the discrepancy, and alerts an editor, ensuring absolute quality control.

Related Reading: To read a complete architectural deep dive explaining how I designed the RAG engine and LLM multi-agent pipelines for this factory, read my post: How I Built the CongKong AI Content Factory.