B10 · Beginner Capstone: AI for Social Good

Building a Prototype

Building a prototype means creating a simple working version of your AI idea to show that it could work and to learn where it fails. For a Beginner capstone, this usually means training a classifier with a no-code tool, testing it on examples it has never seen, recording the results and improving it.

20 min Beginner Jump to the activity

What you need: Nothing. This page starts from the beginning, with no programming and no formulas.

Definition

Team Swachh has 300 labelled photos and a clear plan. Can three students with a school laptop and a web browser build a working classifier in one afternoon? Yes, with a no-code tool. The harder part is testing it honestly.

Building a prototype means creating a simple working version of your AI idea to show that it could work and to learn where it fails. For a Beginner capstone, this usually means training a classifier with a no-code tool, testing it on examples it has never seen, recording the results and improving it.

In simple words

An architect builds a small cardboard model before the real building, to see if the design works and to show others. Your prototype is the cardboard model of your AI idea: small, quick and good enough to learn from.

Why it matters

A prototype turns your idea from a claim into evidence. Its test results, including its failures, are what make your proposal believable, and they show what would need to improve before any real use.

Key terms

Prototype
A simple working version built to test an idea.
Iteration
One round of building, testing and improving. Projects go through several.
Test results
A record of how the model did on examples it had not seen.
Confidence threshold
A level of confidence below which the system says "not sure" instead of guessing.
Failure analysis
Studying the cases where the model was wrong to find patterns and causes.
Minimum viable prototype
The simplest version that can still test the core idea.

Explained in detail

Build the simplest thing first

Team Swachh did not try to build an app, a map and a notification system. Their minimum viable prototype was just the classifier: give it a photo of a bin, and it says "overflowing" or "fine" with a confidence score. If the classifier could not work, nothing else would matter. They created two classes in a no-code tool, uploaded 120 training photos per class, kept 60 aside for testing, and pressed train. It took about a minute.

Test honestly

Truly overflowingTruly fine
Model said overflowing265 (false alarms)
Model said fine4 (missed)25

On the 60 test photos, the first version got 51 right: 85 per cent. But the team looked deeper. Of the 30 overflowing bins, it missed 4. They decided misses were worse than false alarms, because a missed bin stays overflowing for days, while a false alarm only costs one extra stop. Failure analysis showed three of the four misses were evening photos, where garbage on the ground was in shadow.

Iterate

For the second iteration, they added 40 evening photos of each class and retrained. On a fresh test set, misses fell to 1 in 30 and overall accuracy rose to 90 per cent. They also added a confidence threshold: if the model is less than 70 per cent sure, it says "not sure, please send a clearer photo". In their journal, they recorded every iteration: what changed, the test results, and what they learned. They also noted honestly what they had not tested: rainy-season photos and bins in other parts of town.

Showing the whole idea

  • A prototype can include simple mock-ups of the rest of the system, such as a paper sketch of the supervisor's daily map.
  • Show the mock-ups to real users and note their reactions.
  • Be clear about which parts work and which are only imagined.

Finally, they showed the classifier and a paper sketch of the daily map to the sanitation supervisor. She liked the map but asked for one change: list bins in the order a truck would pass them, not alphabetically. That small piece of feedback, which no test set could have revealed, went straight into their proposal. A prototype is as much a tool for conversations with users as it is a technical test.

How it works

  1. Decide your minimum viable prototype: the core thing that must work.
  2. Train a model with a no-code tool using your planned data.
  3. Test on the set-aside examples and fill in the four outcome boxes.
  4. Study the failures, improve the data and retrain.
  5. Record each iteration and what you have not yet tested.

Try it yourself

Guided prototype in the browser

Work through the checklist. Train, test, find the failures, fix the data and test again. Every test is recorded below.

Two iterations of Team Swachh's classifier

The team trains and tests their first model.

Common mistakes

  • Trying to build the whole system instead of the core idea first.
  • Reporting only accuracy and not which mistakes were made.
  • Hiding what was not tested.

Check yourself

Question 1

1. What is a minimum viable prototype?

Question 2

2. Team Swachh's misses were mostly evening photos. What did they do?

Question 3

3. Why add a confidence threshold?

Summary

  • Build the simplest prototype that tests your core idea, using a no-code tool.
  • Test on unseen examples, count each kind of mistake and study the failures.
  • Iterate, record every version and be honest about what is untested.

Finished reading?

Sources: Teachable Machine (Google) · CBSE Artificial Intelligence curriculum resources