Final project · Train a Smart AI Model
Build it. Test it. Explain it.
Course overview · Final chapter
Your brief
A custom-trained AI vision and sound controller with a tested model card.
Train your own custom vision and audio models, diagnose and repair algorithmic blind spots, connect your model to the interactive web runner, and document your findings with an AI Model Card.
Before you begin
Prepare Google Teachable Machine (teachablemachine.withgoogle.com) in Google Chrome, Edge, or Safari before class. Confirm webcam and microphone permissions are granted. Verify the student can save a local .tm project file using "Save project to file" and reload it using "Open project from file". No student Google account or login is required.
Acceptance checklist
- A 3-class visual model (two distinct gestures plus a resting Neutral state) trained and tested.
- A stress-test report testing the vision model against distance, angle, alternate hand, and lighting.
- An audio model with 20 seconds of calibrated Background Noise and two distinct sound/speech triggers.
- A confuser word audit demonstrating that the audio model ignores similar rhymes and ambient chatter.
- An exported model connected to the interactive web runner with tuned confidence threshold and debounce.
- A completed AI Model Card documenting purpose, training data, optimal conditions, and known blind spots.
Keep a test record
For each check, record the action or input, expected result, actual result, and any repair. Repeat the same test after the repair.
| Action/input | Expected | Actual | Change and retest |
|---|---|---|---|
| Choose one test from the Stress Test Matrix | Describe intended classification | Record observed confidence % | Explain dataset adjustment |
Your project tour
- Introduce your model and state its intended purpose.
- Demonstrate the interactive controller running in the web runner with live webcam or audio.
- Explain how training data affects confidence and explain one spurious correlation you fixed.
- Share your completed AI Model Card and identify one remaining limitation.
- Answer one audience question about how training data creates AI bias.
Save your final .tm files and keep your project tour notes. Public account sharing is not required.
Sessions, chapters, and optional projects
The course includes six one-hour sessions covering the core learning outcomes. Chapters are learning units, not a timetable: a chapter may span sessions, and a session may cover parts of several chapters. Optional extension projects may be introduced when students demonstrate the core skills and scheduled time remains. Extension coverage or completion is not included in the course commitment.