Pull-Based Engineering Education at Scale Using Agentic AI
ISE 754 Logistics Engineering, rebuilt for Fall 2026
Friday, September 25, 2026
Computational Friction
It was so difficult to handle the teaching of the software stack to get things done that it overwhelmed all the other issues.
Pull-Based and Flipped: Pulling Before Class
Why pull versus push-based education: Explanation is pulled at the moment of difficulty rather than pushed on a schedule.
The Concept Level: Where the Human Works
Everything covered in logistics engineering relates to developing a model of the logistics system, and so teaching students modeling has always been the core educational goal for these courses. From Lecture 1.2: “Getting through this course was once like building an airplane, with the whole semester spent getting it off the ground. An AI assistant such as Claude Code now makes it airborne almost at once, leaving time to check that it is airworthy, see how well it flies, and even put it through a few loops.”
Concept: the human’s input, Lecture 2.2
→ Formulation: the AI
→ Implementation: the AI
From Recordings to Lectures
What I said: ISE 453 class recording, 15 Sept 2025
The reason for it is nowadays, with everything being online, it doesn’t really matter. But up until about 30 years ago, everything was done actually in printed tariff tables. So it turns out by using 100 weights, the digits were probably the minimum. If you were using tons, then the charge would have, you know, thousands of dollars to ship a ton, and if you did it in pounds, the cost would be less than a dollar. So you’d have a lot of very small fractional numbers. So by doing 100 weight, you ended up with numbers that were in the, you know, you know, tens to 100 range roughly. So it really was, you know, space efficiency when you’re actually printing out the tables.
The AI’s wording over it, in Lecture 3.2, until I approved it
Hundredweight looks like an arbitrary unit to choose, and the reason for it is plain. Until about thirty years ago these tables were printed, and hundredweight is the unit that keeps the printed digits to a minimum. In dollars per ton the entries would run to thousands; in dollars per pound they would be fractions of a dollar needing several decimal places. Per hundredweight they sit in the tens to hundreds, which is the shortest they can be written. The unit is typographic economy rather than anything about freight.
Rules, Regs and Exemplars
A rule is a Python gate. This one, PROSE-13, keeps a developer note such as TODO off every page students read: it scans each line, a hit is an ERROR, and the nonzero exit refuses the commit.
DEV_NOTE = re.compile(
r"\b(TODO|FIXME|XXX|HACK|TEMP HACK|REMAINING TODO)\b")
hit = DEV_NOTE.search(raw)
if hit:
out.append(F(doc, i, "PROSE-13", "ERROR", ...))
fail = n_err > 0 or (args.strict and n_warn > 0)
return 1 if fail else 0An exception, in the rules registry. The figure-reference rule came after Lectures 1.1 and 1.2 were delivered, so both are exempt; the record says who decided it and when, and the rule no longer reports them.
check_figure_refs.py
all except 1.1, 1.2
DECIDED 2026-08-22
by the instructorOne Source of Truth, in Quarto
Lecture 1.1, the source: a hidden cell, then text calling its values
```{julia}
#| echo: false
⋮
fuel_sum = sum(fuel_gal)
fuel_tbar = sum(fuel_thr) / length(fuel_thr)
fuel_avg = length(fuel_thr) * fuel_g(fuel_tbar)
⋮
```
Summing the five trips (@tbl-fuel) gives
`{julia} @sprintf("%.3f", fuel_sum)` gal,
but using the average travel time (…) gives
only `{julia} @sprintf("%.3f", fuel_avg)` gal, …Lecture 1.3, the source: a visible cell, a hidden cell that draws the figure, then text
```{julia}
# Code block 2: simulated cycle time against the VUT formula
⋮
```
```{julia}
#| label: fig-queue
#| echo: false
⋮
lines!(ax, idx, run[idx]; color = :navy)
hlines!(ax, vut; color = :crimson, linestyle = :dash)
fig
```
**One run is one draw.** The convergence in @fig-queue is a
single *run*: one seed, one stream of random draws, …What students read
What students read: the visible cell
and the hidden one’s figure
The Course Website
The home page
mgkay.github.io/ise754f26
Lectures: each one’s page, script, video and twin
A lecture’s page, 2.4: its sections, models and examples at the right
An Overview Video for Every Lecture
Lecture 3.2’s overview on YouTube, 24:44, recorded over the published page: one of eleven so far
Class Settings
Inverse Classroom
- Allows students to learn more about breakpoints they experience during review
- Less discussion about topics already understood
- Time to discuss extensions and validation strategies
GitHub: Course Engine
GitHub
- Allows for all changes to be dated, and committed
- Provides repositories (sharable folders) multiple individuals can write to with access rules
Diff Feature
- Gives users information about what has changed, line-by-line, with each version
Versions
- Any earlier state of content (student or instructor) exists
- Commit histories show (1) how work was done and (2) when it was done
Push/Pull Structure
- Instructors edit and push content, students pull documents and push submissions
- Students pull new material from GitHub
Repositories
- Read-only in —
materials(lectures and scripts) andhandouts(skills, briefs, sheets published) - Read-write out —
work(one private repository for each student) andfeedback(previously graded content)
Commands
- Auto-updating commands housed in GitHub help students interact with all course content
Review: Basis for Pull-Based Learning Inverse Classroom
Course Infrastructure
Validation Classification and Applications
- Screens are meant to ensure that results are plausible
- Expectations help establish reasonable answers based on expectations, similar solved cases, and sensitivity analysis
- Confirmations confirm that a plausible and reasonable result is correct
- Validation checks exist for student understanding and student-AI interactions
- Checks exist to incentivize students to focus on “big picture”, not passively accepting results
- PRIMARY VALUE - these classifications are not the only classifications that could possibly exist, but they capture the different strategies students can use to validate AI modeling and corresponding results
- PRIMARY VALUE - future engineers must be able to effectively assess AI outputs, necessitating this organized classification approach for the course
Client-Consultant Project Functionality and Personal Student Profiles
How the Client Responds
Translatability - Why Do You Care?
Steps of implementation into a classroom
- Build it for ISE 754 functionality
- Generalize a toolkit to maintain specific features, rules, regulations, etc.
- Provide information to users to take generalized toolkit and cater it to new users
- Assess performance and iterate
Questions
Michael G. Kay · kay@ncsu.edu · NC State, ISE
Jake Benhart · jebenhar@ncsu.edu · NC State, OR
Course Lectures
PBIC Toolkit
Research Amp Toolkit for AI-Assisted Research