Skip to content

Case study · The build record

AES 2026 AI Camp

· Tahir Lone

Role
Design, build, instruction
Venue
Wright State University
Date
July 17, 2026
Stack
Next.js, three.js, PostgreSQL, in-browser ML
15
students
3
ML missions
14
cars in the finale
992
top score of 1000
Risograph poster of the Grand Prix podium under a rising sun
Enter the dojo

This is the engineering record of the AES 2026 AI Camp: the brief as it was set, the constraints that shaped the build, the systems that ran the day, and the decisions that kept everything running after the room emptied.

The brief

Fifteen students, one afternoon, zero code, and a hard success criterion fixed before any software existed: every student personally trains working machine-learning models and watches them perform. Anything that could be mistaken for a lecture was out of scope by definition.

Constraints

Everything had to run in the browser on unmodified school lab machines: no installs, no accounts, no GPU assumptions. The room would hold mixed experience levels, so every mission had to start from zero and stay winnable for a first-timer. And because the users were teenagers, the systems had to absorb chaos by design: a scored-attempt budget of three per mission that makes tries precious, pacing the room's leader controls, and progress that survives a refresh.

The design: an arcade, not a curriculum

The camp shipped as a single URL holding three scored missions, each one a different branch of machine learning in playable form: supervised learning by example in Doodle Dojo, classification by judgment in Inbox Defender, and learning by demonstration in the AI Grand Prix, which turns recorded laps into training data. A fourth, unscored playground covers unsupervised learning through clustering. Scoring is tuned toward effort and teaching quality rather than prior skill, which is what makes the leaderboard safe for beginners.

The machinery

Around the games sits the event layer: a leader console that runs the show from lobby to finale, a projector board built for the front of the room, a persistent leaderboard, and a demo mode for rehearsing the whole event in advance. Destructive controls sit behind type-to-confirm. On the day, the software was the co-instructor: the human teaches the room while the console paces it.

One URL, three missions, and a room the software had to survive.

Built to outlive the day

Most workshop software dies at cleanup. This one was designed with an afterlife: a single configuration flip moved the arcade into open play, where anyone can join with just a name. The finale became the living Grand Prix, with the fourteen camp cars standing as a permanent baseline and every visitor's car racing alongside them for a seven-day window. Never resetting the leaderboard is a product decision: the day's record is part of the product.

Nothing about the build is specific to one school or one afternoon. Same URL, new room, new grid.