How AI Apps Work

Step 1: Every AI app has three parts

Think about a translation app. You type a sentence. Something figures out the meaning and matches it with another language. The new sentence appears on your screen.

Almost every AI app follows the same shape:

1. Input: what the user gives the app (text, a photo, a voice recording).
2. Model: the AI that turns the input into something useful.
3. Output: what the app shows or does with the result.

The model is only one part. The app is everything around it: the screen, the code that connects the parts, and the checks that make the result safe and useful. That is the part you will build.

Step 2: A pretend model

In this workshop we’ll start by writing a pretend model, a simple Python function that stands in for a real one. It lets us build and test the whole app without any special accounts.

def ask_model(prompt):
    # A real model would think about the prompt.
    # Our pretend model just returns a fixed reply.
    return "I am a pretend model. You asked: " + prompt

user_input = "What is photosynthesis?"
answer = ask_model(user_input)
print(answer)

I am a pretend model. You asked: What is photosynthesis?

Later we will swap ask_model for a real one. Because the rest of the app only calls ask_model(prompt), nothing else has to change. That is good design.

Step 3: What is an API?

How does your app talk to a model that runs on a different computer? Through an API.

Think of a restaurant. You (the app) don’t walk into the kitchen. You tell the waiter (the API) what you want, using the menu. The waiter brings your request to the kitchen (the model) and brings back the food (the answer).

In code, an API request is usually:

1. You send a request: a message that says what you want.
2. The service sends back a response: the answer, often as structured data.

Here is what a request and response might look like as Python dictionaries:

request = {"prompt": "Explain gravity to a 9-year-old"}
response = {"text": "Gravity is the invisible pull that keeps us on the ground."}

print(response["text"])

Gravity is the invisible pull that keeps us on the ground.

Step 4: Strengths and limits

Modern language models are good at:

• Explaining things in simple words.
• Rewriting, summarizing and translating text.
• Brainstorming ideas.

They also get things wrong:

• They can hallucinate: state false things with total confidence.
• They can reflect bias from the data they learned from.
• They don’t know about events after their training, and may not know your local context.

So, a good AI app never treats the model’s answer as automatically true. It helps the user check it, and it is honest that an AI made it.

Try it: Ask any AI chat tool a question about something you know very well, such as your school or your own village or neighborhood. Did it get everything right?

Step 5: Plan before you build

Before writing code, write a plan. Answer these questions on one page:

1. Who is the app for?
2. What problem does it solve?
3. Input → Model → Output. What goes in and what comes out?
4. What could go wrong? And how will you check?
5. What will the screen look like?

Good ideas for a first AI app are small and specific, like “a helper that explains a science word in simple language”, or “a bot that quizzes me on my vocabulary list”.

For the rest of this workshop you will build your app step by step.