Refract

Drop a photo into a repo — AI critics from three different models tear it apart, an editor fixes it, and the result publishes itself

Refract

A self-correcting photography log

Push a photo to GitHub. That’s it. Refract takes it from there.

A GitHub Actions workflow triggers a three-stage pipeline: multiple LLMs (Claude, GPT-4o, Gemini) analyze the photo and generate critique + improvement suggestions, an AI editor applies the enhancements, and a static site generator publishes the before/after with full documentation of what changed and why.

Pipeline

  1. Multi-Critic — Claude + GPT-4o + Gemini each review the photo independently
  2. Editor — Gemini image generation applies the agreed improvements
  3. Generator — Jinja2 rebuilds the static site; GitHub Pages deploys it

Stack

Python · GitHub Actions · Anthropic Claude · OpenAI GPT-4o · Google Gemini · Jinja2 · PIL