Hey everyone! I'm Ritu, founder of Miniclay AI. We built this because we kept hitting the same brutal wall in our own AI projects: synthetic data looked perfect on paper, but after fine-tuning, the models turned repetitive, lost reasoning, and straight-up got worse. Weeks of compute time and money — gone. The problem we were trying to solve: most tools only catch collapse after training (or even after deployment), which is way too late and insanely expensive. We wanted to fix it upstream — in the data itself — before a single GPU hour is wasted. The approach started simple: basic entropy + similarity checks to detect risk. Then it evolved — we added an auto-rescue loop that calls strong models to generate diverse, safe replacements until the dataset actually passes. What began as a personal frustration hack became a full tool we now want to share with anyone fine-tuning LLMs on synthetic data. Super excited to finally launch it publicly. Would love your thoughts — especially if you've fought synthetic data collapse yourself. What problems are you seeing? Let's talk! 🚀

Miniclay AI

Protect Your AI Models from Collapse

Protect Your AI Models from Collapse