I’m currently building a machine learning system from scratch.
The project trains an AI to solve a geometric packing problem:
how to place cookie-shaped objects efficiently in a limited space.
Key ideas:
• GPT-style architecture for sequence decision making
• Fourier descriptors for shape encoding
• A/B comparison learning (no explicit labels)
• Custom ML pipeline: “TsuMeTsuMe”
The system is written with PyTorch + CUDA and designed to export to ONNX for mobile deployment.
This video shows the training visualization.https://x.com/Masarina002/status/2032972152732004868/video/1
If you like independent AI research, your support helps cover GPU compute ☕