VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
u_Unrolled

About

About Unrolled

First-principles engineering for the frontier of AI.

FREE COMMUNITY THINGS BUT PAID HARD WORK.

Most AI courses fall into one of two traps: either purely academic proofs with no production code, or superficial library tutorials that call high-level APIs without explaining what happens underneath.

Unrolled is designed to bridge that gap. We take you from absolute fundamentals—how CPython manages memory, tensor strides, matrix calculus, and autograd—up to frontier systems engineering: custom CUDA/Triton kernels, PagedAttention, KV-cache hierarchies, 3D distributed training, and Staff-level system design.

The centerpiece of the academy is Build Atlas: a single, living compound AI system built step-by-step from raw arrays of numbers to a custom GPT you train yourself, followed by chat fine-tuning (LoRA), hybrid RAG, autonomous tool-use agents, and continuous-batch serving.

Unrolled is an open-access, independent engineering initiative built by an Anonymous User.

Editorial & Engineering Principles

Every number is measured.
When a lesson says something is faster, smaller or more accurate, it shows the experiment and the result, including when the result is disappointing.
Claims point to sources.
Mechanisms link to the paper, documentation or source code they come from, so you can inspect and verify them.
One coherent project, in order.
Build Atlas teaches each idea at the exact moment the system requires it. Nothing is listed until it is written, tested, and benchmarked.
Zero hype, real engineering.
No invented testimonials, inflated claims, or marketing fluff. If a framework or architectural pattern is discussed, the lesson explains its mechanics and trade-offs in depth.
Changes are public.
Everything that ships is documented in the changelog with reproducible verification scripts.