Descriptions of what an AI agent can do are cheap. This page is the other thing: working software you can open, click, break and read the source of.
The rule for this page. Every line of code here was written by FreeCoder. No human wrote it, fixed it, or tidied it up. A person tested each one and reported the numbers back — the way you would with any engineer — and the agent did the fixing. The original instruction is printed beside each project, along with an honest account of how many rounds it actually took.
A neural network that trains in front of you, in your browser. Watch the decision boundary bend around the data as it learns. Written from scratch — no machine-learning library, no framework, one file.
Build a single self-contained HTML file. It is a LIVE NEURAL NETWORK PLAYGROUND that trains in the browser in real time. Hard rules: one file only, no external libraries, no CDN links, no build step, everything inline. 1. A canvas showing 2D training points in two classes, with a selector for three patterns: spiral, two circles, and XOR. 2. A feed-forward neural network written from scratch in plain JavaScript (no ML library) with a configurable hidden layer. 3. Train it live with backpropagation and animate the decision boundary repainting as it learns. 4. A second small canvas plotting the loss curve updating in real time. 5. Controls for pattern, learning rate, hidden layer size, and Train / Pause / Reset. 6. Dark theme, background #0b0b12, violet accent #7c3aed, polished and professional. When you are done you MUST verify it actually works... Fix anything broken and re-check before you tell me it is finished.
The first two attempts produced a beautiful interface attached to a network that did not
learn — 64% accuracy, then 49%, which is guessing. The agent could not tell,
because it cannot open a page and look at it, and broken maths and working maths read
identically in source.
So it was asked to prove the maths instead of inspecting it. It pulled the network out
into a separate file with no interface code, wrote a test that trains on XOR and on the
spiral and prints the loss as it goes, and ran it. Both passed. Then the page worked.
An agent that writes a test to check itself is doing the job properly.
That is the interesting result here — more than the demo.
New projects are added as the agent builds them. Each one arrives the same way — one instruction, no human edits.
irm https://freecoder.space/install.ps1 | iex — then read the
getting-started guide.