Technical Curriculum

Teaching Model Evaluation to Developers

A five-minute video lesson on evaluating a trained logistic regression model — storyboarded, scripted, and produced for software developers who can code but are new to machine learning.

The finished lesson — 4 min 19 sec. The storyboard and full script are linked below.

The Problem

Machine learning material written for developers usually fails in one of two directions. It either assumes the reader already thinks in linear algebra, or it strips out so much that nothing transfers to real work. The audience I designed for sits precisely in between: software developers who are comfortable with code, math, and abstraction, but who have never trained a model.

The brief was narrow on purpose — one concept, five minutes. Teach how to evaluate a logistic regression model that has already been trained, and stop there.

The Design Decision That Mattered

The source material framed model testing as four numbered steps. Following that structure would have been the fast path, and it would have produced a worse lesson.

The first of those four “steps” was not a step at all — it was the inputs: the feature matrix, the labels, and the learned parameters. Those are setup, not action. Presenting them as step one meant a learner had to hold four things in working memory when only three of them were things they would ever do.

So I restructured it. Inputs became context established up front. The lesson itself was built around three computational steps — probabilities, then predictions, then accuracy — each producing one named output. Three named outputs is a load a learner can carry through a five-minute explanation, and it let the closing recap mirror the opening preview word for word.

That is the whole job in miniature: subject-matter accuracy is necessary but not sufficient. Someone has to decide what shape the knowledge takes before a learner meets it.

Scope, and What I Deliberately Left Out

Five minutes buys you one idea, taught properly. Everything else has to go somewhere else in the sequence.

In scope

•The three inputs and how they relate to held-out data
•Probabilities → predictions → accuracy
•Reading an accuracy number in context, via two contrasting domains
•How that number points to what you change next

Deliberately deferred

•How the model was trained
•The mechanics of gradient descent
•Fixing overfitting — named, then handed to a later lesson

Production

The deliverables were an eight-slide storyboard with full speaker notes, and a produced video under five minutes.

The automated slide-to-video conversion dropped the animations, which sounds like a formatting problem and is actually a pedagogical one. During a mathematical explanation, what the learner sees and precisely when they see it is the instruction. I rebuilt the highlights and reveals by hand so each term appeared as it was spoken, not before. Manual video editing was new to me on this project; the decisions driving it were not.

See the storyboard

All eight slides with the full speaker notes — the script exactly as it was delivered.

Open the storyboard (PDF)

Working With AI

I used AI heavily and as a collaborator rather than an author. Claude Code drafted and iterated the script; I pushed back whenever a proposed structure conflicted with what I knew about cognitive load — the four-steps-into-three restructure came out of exactly that argument. For the visuals, when generated layouts didn't give me the control I needed, I built HTML mockups and converted those into slides instead. Slower than accepting the first output; faster than starting from nothing.

To make sure the technical content was correct, I used AI as a tutor before using it as a tool — working through logistic regression with adaptive questioning until I could defend every claim in the script myself. Every instructional and editorial decision stayed with me. That is the distinction I care about: AI compresses production time enormously, and it has no opinion about what a learner can hold in their head.

Project Details

Role: Instructional Designer & Producer
Audience: Software developers new to ML
Tools: Claude CodeHTMLGoogle SlidesVideo editing
Category: Technical Curriculum / Video
Year: 2026