25% understand. 75% build. Learn just enough of what’s under the hood to never be fooled — then spend the rest shipping real AI products in a STEM or business field you care about, with guardrails that keep it honest.
See how it works →The fastest way to lose a class is “math first, cool stuff never.” The fastest way to create fragile builders is “ship magic you can’t explain.” This track threads the needle: a short, vivid foundation that kills the magic and installs a BS-detector, then four weeks of building real things with AI — where you must measure, break, safeguard, and defend everything you ship.
The grounded 25% isn’t a lecture — it’s the smallest set of hands-on truths that make you impossible to fool: features/label/loss, the train/test split, bias you can measure, and “plausible ≠ true.” Then you spend the 75% where the motivation (and the resume) actually is: building.
Both are great. Pick by where you’re headed.
One fast, hands-on session: train a tiny model, measure it honestly, produce & catch bias by hand, and see why fluent AI can be confidently wrong. You leave with a BS-detector — enough to never be fooled by an AI, or by your own demo.
Explore how AI is remaking real fields (see AI across STEM & Business →), choose a domain and a real user with a real problem, and scope your build with the canvas. Data + rights checked up front.
“Vibe coding” done right: describe what you want, let AI draft the code, run it, read the errors, iterate. Ship a working prototype end-to-end. Ugly is fine; running is required.
This is where building earns its rigor: measure a real number, red-team your own product until it breaks, add one safeguard, and be able to explain every part. “The AI wrote it” is not an answer.
A public 3-minute, 4-beat demo with one real artifact and your number said out loud. Walk away with a shipped product, a resume bullet, and a certificate.
Building fast with AI is a gift — and a trap. These five non-negotiables import the discipline of a real engineer into the speed of a vibe coder. Every capstone must pass all five.
Accuracy, recall, pass-rate, a user metric — something on a held-out test, not a vibe. If you can’t measure it, you can’t claim it.
A pod-mate tries to break your product — wrong inputs, edge cases, prompt injection. You fix what they find.
A threshold, a refusal, a dry-run confirmation, a “not medical advice” note. Name the harm you anticipated and the guard you built against it.
You must be able to explain what each piece does — even the code AI wrote. If you can’t explain it, you don’t ship it.
Where did the data come from? Do you have the right to use it? Whose privacy is involved? Answer before you build.
Leverage is only yours if you can stand behind what it produced.
Wk 1 — trained & measured a model, caught bias by hand, can spot a confident-but-wrong AI.
Wk 2 — picked a real arena, a real user, and scoped a build with data rights checked.
Wk 3 — shipped a working MVP end-to-end using AI, and can explain how it works.
Wk 4 — measured a number, red-teamed it, and built a safeguard that works.
Wk 5 — public demo with one artifact + a resume bullet with a real number.
All 5 badges + all 5 guardrails passed on your capstone.
Claude, ChatGPT, Gemini — your build partner and explain-only tutor.
Replit, Lovable, Cursor, v0, Claude Artifacts — describe it, ship it.
Just for Week 1’s grounded foundation — enough Python to not be fooled.
Turn your build into a public web demo people can try.
Learn to red-team — break AI safely before you defend your own.
Whatever your field needs — librosa (audio), MediaPipe (pose), public datasets.