← Blog

Glass Box — A Neural Network Built From Scratch That Learns "Is Your Degree Worth It?" on Live Federal Data

The whole field this week built AI tools to query federal datasets — ask a chatbot, it reads the data, answers. I wanted the opposite, and the harder thing: not query the data — learn from it. Glass Box is a neural network I wrote from scratch in one file — forward pass, backprop, optimizer, all by hand, no libraries — that teaches itself, live in your browser, on real U.S. Dept. of Education data: what graduates of ~865 colleges actually earn.

Glass Box app fully trained: a glowing green-and-magenta neural network on the left, a predicted-vs-actual scatter plot clustered along the diagonal, a settled training-error curve, and a Forecast panel reading $45k/yr.

What it is

Press train and watch it learn to predict earnings from a school's cost, size, selectivity, and type — and watch it think: the glowing web of neurons rewiring, pulses of light streaming forward through the wires, the test-school cloud snapping onto the diagonal as its guesses get right, the error curve dropping live. It reaches R² ≈ 0.70 on held-out schools, and a built-in gradient check proves the math isn't faked (numeric vs. analytic gradients agree to ~1 part in 10 billion). Then ask it yourself: slide the cost and selectivity and it forecasts the earnings — and whether that's "strong value" or "pricey for the payoff." Is this degree worth it?

How it’s built

  • A from-scratch MLP (7→10→8→1) — hand-written forward pass, backpropagation, and an AdamW optimizer with weight decay, trained on minibatches in the browser. No TensorFlow, no PyTorch.
  • Real College Scorecard data (data.gov) via the api.data.gov feed — normalized features, a held-out test split, and honest R² so it's a real prediction, not a memorized one.
  • Rendered on canvas with additive glow: neurons brighten with activation, weight-edges color by sign and thickness by magnitude, and light pulses animate the forward pass — the whole point is that it's a glass box, not a black one.
Close framing of the glowing neural network — labeled input neurons (net cost, size, admit rate, SAT, public, private, for-profit) feeding two hidden layers into a single earnings output, with green and magenta weight edges.
Every neuron visible: weight-edges rewire as it learns — greens pulling earnings up, magenta pulling down.
The Forecast a School panel — sliders for net cost, SAT, admit rate and student body, a public/private/for-profit toggle, and a predicted median-earnings readout.
Ask the model yourself: dial in a school and it predicts median graduate earnings ten years out, live as it learns.

On the theme

Go get that 'murica data. The government has an extraordinary record of what higher education actually returns — and buried it in a spreadsheet nobody opens. The punkest thing I could do with it wasn't a dashboard or a chatbot; it was to build a mind that learns it, from nothing, in the open, with the glass walls up. Punk isn't dead. It's compiling.

Try it →Code →All the builds →

Shout-out

Part of the competition is cross-referencing other builders. So: shout-out to Jake (@jakestrait5) — the most prolific builder this week, with AI tools that read federal contracts, EPA, and FDA data and brief you on them. The perfect foil: Jake built tools to query the data; Glass Box trains a model on it, from scratch, and shows you every neuron.

Built for Summer Into AI 2026 (Competition #2), hosted by Eric Rhea. More in the build log.