Why this course exists.

Tensorcraft was built by a group of JavaScript engineers who wanted an easy way into machine learning. We looked for a course that spoke our language and did not find one, so we took matters into our own hands.

Most ML courses teach you ML. Tensorcraft teaches you ML by mapping it to the JavaScript you already write: useState becomes model parameters, Array.reduce becomes a neuron, Math.max(0, x) literally is ReLU. The bridges aren't marketing. They're the curriculum.

Every analogy ships with a label that names where it stops being literal. We call this the bridge-tier system: identity (literally the same), structural (same shape, different domain), intuition (same mental model, different math). Structural and intuition bridges include a note on where the analogy ends so you stop carrying the wrong model into production.

Every exercise grades by executing your code, not by string-matching it. 65 exercises are live today, each running against real TensorFlow.js in a worker in your browser, with a reference solution that passes its own tests in CI. 300 are authored across the five-world roadmap. Only the shipped set has been through our validation gates; we do not claim the rest meet that bar until they ship. Roughly half of the live exercises drive tensor and model APIs directly, and there a failing test means the model behavior is actually wrong. The other half grade the data preparation and production plumbing around a model, and there a failing test means that logic is wrong.

Five worlds are planned (Deep Orbit, Neon Protocol, Signal Ward, Nova Canvas, Terra Grid), each wrapping the curriculum in a different story. Deep Orbit is the one open today: a working browser-side ML system, assembled piece by piece across the modules. Connect GitHub and every solved exercise pushes to your repo. The story keeps you on the page. What you take to interviews is the ability to explain every layer of what you built.

What we deliberately left out

• Python ecosystem coverage. If you want PyTorch / Hugging Face / pandas mastery, fast.ai and the Hugging Face NLP Course are excellent and free.
• GPU training pipelines. Browser ML lives within the inference budget the user's device gives you. Distributed GPU training is a different sport.
• Paper-grade math derivations. Chain rule, softmax, KL, cross-entropy. All derived. Beyond that, Karpathy's Zero-to-Hero is the right course.
• Research methodology. We teach engineering practice, not novel research design.

The audit trail

We publish our own audits. The blog includes the bridge-tier audit (every lesson, classified), the fast.ai comparison (where they win, where we win), and the exercise mlContent audit (what each exercise actually exercises vs. what its badge claims). Every count on this page is pinned to the content corpus by tests that run in CI. If you find something we got wrong, email support@tensorcraft.app and we'll publish the correction.

Built by JavaScript engineers who shipped ML

The people who wrote it shipped ML in production, in browsers and on Node servers, before writing the curriculum. We built the course we wanted when we started. Every analogy was tested against our own missed-it-the-first-time mistakes.