ClearDataLabs is an open-source educational project that explains AI concepts by building neural networks from scratch — no TensorFlow, no PyTorch, just TypeScript and a browser. Every article includes working code, and every demo runs client-side with no server.

Most machine learning tutorials hand you a library and a dataset. You call model.fit(), watch the accuracy climb, and walk away with a trained model and no particular understanding of what just happened.

That’s useful. But it’s not the only way to learn.

ClearDataLabs exists for the other approach: understand what’s actually going on. How do networks learn? How does data flow through layers? What does a trained model really “know”? These questions don’t have one-line answers — but they become a lot clearer when you can see the mechanics, interact with them, and sometimes build them yourself.

Why fundamentals matter

AI is moving fast. New architectures, new training techniques, new applications — the landscape shifts constantly. But the core ideas behind all of it — gradient descent, loss functions, representation learning, optimization — remain remarkably stable. Understanding those fundamentals makes it easier to follow what’s new, evaluate what’s hype, and build on what actually works.

ClearDataLabs focuses on those fundamentals. Not to ignore the cutting edge, but because a solid foundation is the best tool for navigating it.

The projects so far

Handwriting Recognition in the Browser — a feedforward neural network that reads handwritten characters in real time. You draw a letter, the network classifies it. Trained and running entirely in the browser, with backpropagation implemented from first principles.

Knowledge Extraction from a Neural Network — an MNIST digit classifier, trained and frozen, dissected two ways. First, by computing a causal index that measures how much each input pixel influenced each output class. Second, by optimizing a blank image until the network “sees” a specific digit in it — a browser-native version of DeepDream.

These are starting points. The project is a playground for experimenting with different architectures, training methods, and visualization techniques — open-source and not tied to any specific framework or brand.

What you’ll find here

Articles that go deep on the ideas and their implementation. Not “here’s the concept,” but “here’s how it works, here’s the code, here’s where it breaks.” Written for anyone curious enough to want to understand AI, not just use it.

Start with the projects page for the demos, or jump straight into the articles.