This project builds an end-to-end automated content publishing workflow to reduce the time gap between international technical sources and Traditional Chinese readers.

📋 Abstract

Technical news curation often depends on manual translation and reposting, which can introduce delays of days or weeks. To reduce that cost, I set up an n8n workflow on a Linux server and integrated Ollama local large language model (gpt-oss) for RSS monitoring, content extraction, summarization, translation, image retrieval, and WordPress publishing. The system is designed to produce reviewable drafts first, leaving final judgment to a human editor when needed.

🛠️ Technical Case Study

1. System Architecture & Infrastructure

To ensure system stability and privacy security, this project adopted a fully privatized deployment solution:

  • Linux Server: Serves as the foundational environment for computing and service hosting.
  • Workflow Orchestration (n8n): Uses n8n as the automation hub, responsible for connecting various API nodes and logical judgments, replacing traditional cumbersome Python Crontab scripts.
  • Local LLM Inference (Ollama): Runs the gpt-oss model through Ollama, reducing dependence on external APIs for large-volume text processing.

Lightnews automated content pipeline

2. Core Technology: AI-Driven Content Pipeline

The core of this system lies in transforming unstructured web content into structured publishing format through AI.

2.1 Ingestion & Cognitive Processing

The workflow begins with RSS monitoring of specific technical domains. Once new articles are discovered, the system triggers the following processing chain:

  1. Content Cleaning: Automatically crawls original web pages, removing ads and irrelevant HTML tags.
  2. Local LLM Inference: Feeds cleaned text to Ollama (gpt-oss) for multi-dimensional processing:
    • Summarization: Extracts core technical points from articles.
    • Translation: Accurately converts English technical terminology to Traditional Chinese.
    • Auto-Tagging: Based on semantic context, automatically determines the technical domain (e.g., DevOps, AI, Security) and generates corresponding tags.

2.2 Context-Aware Media Retrieval

To make articles visually engaging, I designed a “text-to-image” retrieval logic rather than simply using random images.

  1. Visual Intent Recognition: LLM analyzes article content to generate searchable English “Visual Keywords.”
  2. API Matching: System automatically calls image library API (Unsplash) using these keywords for search.
  3. Optimized Selection: Based on download count and relevance scores, selects a relevant image as the article’s Featured Image.

3. Automated Delivery

In the final stage, n8n sends processed titles, content, tags, and image links to the WordPress REST API to create drafts or publish posts. This keeps article formatting consistent and supports a steadier update rhythm.

Conclusion

This project demonstrates how Low-Code tools (n8n) and Local LLMs can support a practical content workflow. The goal is not to replace editorial judgment, but to automate repetitive extraction, summarization, and pre-publishing work so human review can focus on quality control.