Read the mechanics.
Trace reverse-mode differentiation, inspect NumPy kernels, and follow the maths into the code.
Read autogradTHE OPEN-SOURCE NEURAL COMPUTING FRAMEWORK
Understand the
intelligence
you
build.
From your first tensor to your own transformer. A readable Python framework built on NumPy, for developers who want to see how it all works.
LIVE CONCEPT / AUTOGRAD
01
FORWARD PASS ↓
INPUTx 2.0
× x · x
OUTPUTy 4.0
∂y / ∂x 4.0← backward()
NumPy at the core.Every operation, readable.
YOUR CURIOSITY.
THE WHOLE STACK.
01 / WHY EVERYO
Less black box.
More lightbulb moments.
Use EveryO to learn neural-network internals, teach with runnable Python, or inspect a small CPU experiment. Follow the implementation from a tensor to a parameter update.
Own the experiment.
Generate datasets locally, set your seeds, and train small models without an account or API key.
Find an experimentCheck the numbers.
Explore finite-difference tests and optional TensorFlow reference comparisons. Understand what was validated.
See validation02 / A FEW LINES. A REAL GRADIENT.
The maths is yours
to follow.
Mark a tensor, perform an operation, and call backward(). EveryO records the graph and accumulates the gradient.
import everyo as eo
x = eo.tensor([2.0], requires_grad=True)
y = x * x
y.backward()
print(x.grad) # [4.]
Expected: [4.] — because d(x²)/dx = 2x
Documented examples. Python runs on your machine, not in this page.
03 / ONE READABLE STACK
Build from the ground up.
The same NumPy foundation takes you from tensor arithmetic to convolution and attention.
04 / EXPERIMENTS YOU CAN REPRODUCE
Show your work.
Explore the training curves and decision boundaries produced by the framework’s own examples.
DOCUMENTED EXAMPLE RUNCPU
99.3 percent held-out digit accuracy
Held-out digit accuracy
- Held-out samples
- 600
- Trainable parameters
- 17,226
An example-run measurement, not a benchmark or performance guarantee. Results depend on hardware and configuration.
Reproduce the digit experiment
05 / MADE FOR THE CURIOUS
Your next “what if?”
belongs here.
Choose a starting point. Make a small change. See what happens.
Follow a gradient.
Start with tensors and understand why the derivative of x² is 2x.
Learn the fundamentalsTrain something small.
Build a classifier, visualize a decision boundary, or inspect an attention head.
Explore examplesMake it better.
Bring a reproducible bug, clearer docs, or a well-tested implementation.
Meet the projectFROM “HOW DOES IT WORK?” TO “I BUILT THIS.”