The project combines Tunix, Gemma and Gemini Flash 3.7 to let agents iterate on their own on model fine-tuning.