Show HN: Syna – Minimal ML and RL Framework Built from Scratch with NumPy (github.com)

Syna is a compact, educational ML and RL framework implemented from scratch on top of NumPy. Announced as a Show HN project, it adopts a define-by-run (dynamic computation graph) design inspired by DeZero and bundles basic reinforcement-learning tools (e.g., a DQN agent + Trainer) in the same library rather than splitting RL into a separate package. Its goal is pedagogical: to expose the inner mechanics of autodiff, graph construction, and training loops in plain Python so beginners and researchers can learn how frameworks like PyTorch work under the hood. Performance and GPU support are intentionally omitted to keep the code short and readable. Technically, Syna supports higher-order gradients via backward(create_graph=True) and provides a utility to visualize computation graphs with Graphviz (example: fifth derivative of tanh(x)). It includes an out-of-the-box DQN example that can solve CartPole-v1 and is installable from GitHub or PyPI (MIT licensed). If you want a minimal, transparent reference implementation to study dynamic autodiff, gradient propagation, and simple RL pipelines, Syna is a tidy, hands-on resource — just expect NumPy-only CPU execution and deliberate simplicity rather than production throughput.