Deep Orbit

Assemble and compile an untrained classifier: a neuron (a weighted sum written with reduce), the activation that keeps stacked layers from collapsing, layers that cost parameters, and the batch axis a reading carries.

Loss functions score each guess, gradient descent nudges the weights to shrink that score, backpropagation gets every nudge from one backward pass, the loop ties them together, and evaluation decides what ships.

Run a trained model in the browser: TensorFlow.js setup, exporting, loading and caching a model, real-time inference (a prediction per reading), and how TensorFlow.js, ONNX Runtime Web, and Transformers.js compare.

Use the Fourier Transform (which frequencies a signal contains), spectrograms (those frequencies pictured over time), and windowing (tapering each slice's edges) to turn raw readings into inputs a model learns from.

Recurrent networks that carry state across time: RNNs and LSTMs mapped onto the state-update code you already write, plus the split-by-time rule that keeps a windowed stream from scoring itself.

A baseline that says what normal is, an autoencoder that learns it, reconstruction error scored against a tuned cutoff, and a 2-of-3 vote across three detectors that can still fail together.

Explore attention, forecasting, Kalman filtering and ensembles. Compare their assumptions, inspect errors and decide what each method can support.

Add language models to the pipeline: prompt engineering, retrieval over your own logs, and orchestrating more than one model.

Shrink a model to fit the hardware: profiling, quantization (fewer bytes per weight), pruning (zeroing small weights and measuring the effect), caching, and WebGPU (the math on the GPU), a trade of size against accuracy.

Assemble the finished system: architecture, a deployment pipeline, monitoring, the documentation that ships with it, and one rehearsal that trains the deploy candidate and gates it.