Gemini AI turns your text prompts into JSON-based shape instructions
The server renders the image using Python PIL on a plain Linux CPU (no GPU, no canvas)
The site is intentionally slow — CPU rendering is inefficient and queue-based
Limited to 2 requests per day due to Gemini API free tier limits
You can watch your request crawl through a real-time processing queue
It’s purposefully “bad” — just basic geometric shapes — but somehow still oddly entertaining.
The whole point was to see if we could build something fun using only CPU rendering and free-tier AI.
Live demo: https://poops-draw.site Tech: FastAPI, Gemini API, PIL, MySQL, Docker