miniGPU 🫖 — A clean GPU workflow, from GLSL all the way to the teapot.

minirt 🔦 — A minimalist ray tracing processor (a sister device to miniGPU)

ToyGPU ⚔️ — The Minimal Complete Compilation Process for Software GPU True LLVM Backend

ClawX is a thin PTY wrapper + soul framework for Claude Code CLI. A set of tiny config files that give Claude Code persistent identity, memory, heartbeat, and scheduled tasks — all within the offic…

A tiny language model trained on Nami's own memory. Pure NumPy via numpy-grad, pure CPU, pure offline.

Array-level reverse-mode autograd in pure NumPy. Same Tensor API as cpp-grad / micrograd, but each node wraps an np.ndarray instead of a single float — so matmul runs on BLAS at full speed and the …

A protocol for an LLM agent to autonomously run ML training experiments overnight. This repo is not training code. It is a spec: the playbook the agent reads before it starts, plus the disciplines …

Here are some tips and notes on programming.

Inspired by karpathy/autoresearch. The idea: give an AI agent a small but real LLM training setup and let it experiment autonomously. It modifies the code, trains for 10 minutes, checks if the resu…

A deep learning framework built from scratch. It implements forward propagation, backpropagation, and gradient updates purely in C++, without relying on PyTorch or TensorFlow. From MNIST CNN to GPT…