LATENTFORM
Initializing weights 000%

Neural engineering × immersive 3D

Intelligence,
hand-built.

I design neural networks from first principles and turn them into real-time, interactive experiences: the math, the model and the motion, under one roof.

[ move your mouse through the core ] Scroll

02 — Manifesto

Most products rent their intelligence. I write mine, from the first matrix multiply to the last pixel on screen. The networks further down this page have no TensorFlow and no PyTorch behind them: just hand-written math, learning live in your browser.

03 — How a model is born

01

Data

Collect & shape the signal.

Every model starts as a cloud of examples. Cleaning, labelling and augmenting that cloud is half the craft. Garbage in, garbage out.

  • Cleaning
  • Labelling
  • Augmentation
02

Architecture

Design the network.

Layers, widths, activations, initialisation. Choosing the shape of the function before it learns anything decides what it can learn.

  • MLP
  • CNN
  • Transformer
03

Training

Descend the loss landscape.

Backpropagation works out how every weight should change, and an optimiser like Adam takes the step. The glowing trail is one descent into a valley.

  • Backprop
  • Adam
  • Regularisation
04

Deploy

Ship it where people are.

In the browser, at the edge, on device. Fast, private inference inside an interface people actually enjoy using.

  • WebGL / WebGPU
  • ONNX
  • Edge

04 — The Models

Three neural networks, written from scratch. Train them live.

No ML libraries. Each model has its forward pass, backpropagation and optimiser hand-written in JavaScript, and trains in your browser right now: one learns to classify, one learns to write, one learns to balance.

Click to add a point · hover to probe
Live network · 2 → 12 → 12 → 3
positivenegative
Epoch0
Loss—
Accuracy—
Params—
Dataset
Click adds
Hidden layers
2
Neurons / layer
12
Speed
Activation
Training curve · loss · accuracy
The actual backward() running on this page

05 — What I build

From the first weight to the final frame.

Neural network engineering

Custom architectures designed, implemented and trained from first principles, or with PyTorch and JAX when the job calls for scale.

  • MLP
  • CNN
  • Transformers
  • Backprop

Real-time 3D & WebGL

Hand-written GLSL, GPU particle systems and post-processing pipelines. The background of this page is one.

  • Three.js
  • GLSL
  • WebGPU
  • Shaders

Interactive storytelling

Scroll-driven stories, physics-based motion and small interactions that make complex ideas feel easy.

  • GSAP
  • Scroll
  • Motion design

On-device AI & deployment

Models that run where your users are: in the browser, at the edge, on device. Private, fast and cheap to serve.

  • ONNX
  • Quantisation
  • Edge
0ML libraries behind the three live models
30kparticles simulated on the GPU, behind this text
—trainable parameters in the lab network, right now
3fields of machine learning, trained live in your browser

06 — Contact

Have a problem worth a neural network?

Start a project betty.ballin@yahoo.de