I’ve been running my own Pleroma instance at social.curlyphries.net for a while now and recently I decided to take it a step further by building an AI-powered bot that lives there. Meet Pixel - a bot that analyzes news and provides intelligent commentary when you mention it.

What is Pixel?

Pixel is a Fediverse bot that combines real-time news data with local AI analysis to provide thoughtful insights on current events. Instead of relying on cloud-based AI services that might compromise privacy or cost a fortune, Pixel runs entirely on my own infrastructure using open-source tools.

How It Works

The architecture is surprisingly straightforward:

  1. Social Platform: Pixel runs as a bot account on my Pleroma instance
  2. News Source: It pulls current events from TheNewsAPI
  3. AI Analysis: Local Ollama instance processes the data and generates responses
  4. Interaction: Users can @ mention Pixel to trigger analysis on specific topics

When you mention Pixel in a post, it:

  • Fetches relevant news articles from TheNewsAPI
  • Processes the information through Ollama (running locally)
  • Generates a thoughtful analysis or summary
  • Replies to your post with the insights

Why Build This?

There are a few reasons I wanted to create Pixel:

Privacy & Control: By running everything on my own infrastructure, I maintain complete control over the data and interactions. No third-party services logging conversations or training on my data.

Learning Experience: Building a bot that integrates multiple APIs and AI models is a great way to understand how these systems work together. It’s one thing to use ChatGPT; it’s another to build your own pipeline.

Fediverse Community: The Fediverse is all about decentralization and self-hosting. Having a bot that can provide useful information while respecting those principles feels like a natural fit.

Cost Efficiency: Running Ollama locally means no per-token API costs. The trade-off is electricity - running the model on my own hardware does consume power - but it’s still far more economical than cloud API pricing and I have complete control over when and how the model runs.

The Tech Stack

Here’s what powers Pixel:

  • Pleroma: The lightweight, efficient Fediverse platform that hosts the bot account
  • TheNewsAPI: Provides access to current news articles and headlines
  • Ollama: Open-source tool for running large language models locally
    • Model: llama3.1:8b (4.92 GB)
    • 8 billion parameter model offering a good balance of quality and speed
    • Handles all bot responses, research analysis and conversations
  • AWS EC2: Cloud infrastructure hosting the Pleroma instance
  • Python: Glue code that ties everything together

What Makes It Interesting

The combination of real-time news data with local AI creates some interesting possibilities:

  • Fact-checking: Cross-reference claims with current news sources
  • Summarization: Get quick summaries of complex news stories
  • Perspective: Analyze how different sources cover the same story
  • Context: Understand the background and implications of current events

Privacy & Security Considerations

Running a bot on a public social platform requires thinking carefully about security:

  • API keys and credentials are properly secured (not hardcoded)
  • The bot runs with minimal permissions
  • Rate limiting prevents abuse
  • The infrastructure is isolated from other services
  • No user data is stored beyond what’s necessary for functionality

Beyond Text: Comic-gen

While Pixel focuses on text-based analysis, I’ve also been experimenting with visual content generation through Comic-gen - a custom web application I built that wraps around the FLUX image generation model.

What is Comic-gen?

  • Custom interface for generating comic-style images
  • Uses FLUX, a state-of-the-art open-source image generation model
  • Known for high-quality, detailed images
  • Works efficiently at low guidance scales (CFG 1.0)
  • Optimized for comic and illustration styles

This isn’t integrated with Pixel yet but it represents another piece of the self-hosted AI puzzle - proving that you can run sophisticated image generation models on your own infrastructure, just like text generation.

Future Plans

I’m constantly thinking about ways to improve Pixel:

  • Adding more news sources for diverse perspectives
  • Implementing conversation memory for multi-turn discussions
  • Creating specialized modes (e.g., tech news, politics, science)
  • Improving response quality through prompt engineering
  • Adding source citations to responses
  • Potentially integrating Comic-gen for visual content in responses

Closing Thoughts

Pixel represents what I love about the Fediverse and self-hosting: the ability to build custom tools that work exactly how you want them to without compromising on privacy or control. It’s not about replacing commercial AI services - it’s about understanding how they work and creating alternatives that align with your values.

The Fediverse is more than just a Twitter alternative. It’s a platform for experimentation, learning and building tools that serve communities rather than shareholders. Pixel is my small contribution to that vision.

Want to learn more about self-hosting Fediverse instances or building your own bots? Feel free to reach out on the Fediverse or check out my other posts on decentralized social media.