I Can Build Apps with AI. But I Realized I Didn't Really Understand How AI Works Inside Them.

Today is hackathon day!

I'm attending Tokyo Agent Hackathon, held on Saturday, October 17, from 9:30 AM to 6:00 PM.

This morning, while getting ready to leave, I was thinking about what kind of app I might want to build.

Over the past few months, I've built quite a few apps using tools like Claude Code and Codex.

So I've gradually become more comfortable with building apps with AI.

But this morning, I realized something.

I can use AI to build an app. But I still don't fully understand how to use AI inside an app I'm building.

Maybe I Could Build an AI Pet That Helps People Reach Their Goals

I spent some time thinking about what would be fun to create at today's hackathon.

My first idea was an AI pet that responds to my daily activities and changes over time, almost like a pet I could keep raising.

Imagine a cute character on your screen that reacts to your actions, habits, and progress.

As things change in your life, your pet changes too.

That could be fun!

Then I started thinking about what the pet could respond to.

Three ideas came to mind.

1. A Pet Connected to Fitness and Body Goals

The first idea was fitness and health.

What if the pet could respond to things like body fat percentage, muscle mass, or calorie intake?

If I maintained my body fat percentage, the pet might celebrate.

If my muscle mass increased, maybe the pet could grow a little too.

It could also react to exercise or food records.

Instead of simply entering numbers into an app, I could see some of those changes reflected in my pet's appearance or movements.

Maybe that would make tracking progress more enjoyable.

2. A Pet Connected to Work and Task Management

The second idea was work.

The pet could respond to completed tasks or progress on a project.

For example, when I finished something on my to-do list, the pet could celebrate with me.

For a bigger project, it could show how far I'd come and how much was left.

Rather than just checking off completed tasks, I might actually feel like I was moving toward a goal together with my pet.

It could make work feel a little more like a game.

3. A Pet Connected to Savings and Money Management

The third idea was personal finance.

I thought about tracking savings goals or the balance between income and expenses.

For example, if I were saving for a trip, the pet could celebrate as I got closer to my target amount.

Instead of being just another budgeting app where I enter income and expenses, it could make progress toward a goal more enjoyable.

But I wouldn't want the pet to become sad or weak just because I spent more money one month.

Money management should feel encouraging too.

Wait, Why Not Let People Set Their Own Goals?

After thinking about fitness, work, and money, another idea occurred to me.

What if I didn't have to choose just one of those themes?

Maybe users could set their own goals, and the AI pet could accompany them as they worked toward those goals.

That way, the same idea could work for different parts of life.

For example, imagine setting a work goal:

Publish three blog posts by the end of October.

After publishing the first post, I could tell my pet:

"I published one blog post today!"

The progress bar would update.

1 out of 3 posts. 33% complete!

And the pet might say:

"Yay! Just two more to go!"

After the second post, the progress would reach about 67%.

And after publishing all three, the pet could jump, dance, sparkle, and celebrate with me!

In a typical task management app, completing something might simply mean checking a box.

But wouldn't it be more fun if a cute pet celebrated with you?

I also realized that goals come in different forms.

Some involve reaching a total, like publishing three blog posts.

Others involve numerical targets, such as body fat percentage or savings.

And some are about consistency, like studying for 30 minutes every day.

If the app could display progress differently depending on the type of goal, it might work for fitness, work, money, studying, hobbies, and all kinds of other activities.

I also wanted users to be able to talk naturally to their pets instead of always entering numbers.

For example, I might say:

"I published a blog post today! It took a lot longer than I expected, though."

The AI could understand that I'd completed one post, ask me to confirm the progress before recording it, and respond to how I was feeling.

Something like:

"That sounds like a lot of work, but you finished it! Great job!"

Then I thought about something else that really mattered to me.

I wouldn't want the pet to become weak, angry, or disappointed when I failed to reach a goal.

Some days we're busy, exhausted, or simply unable to make progress.

If opening the app meant seeing a sad pet or feeling guilty, I might eventually stop wanting to use it.

I'd rather have a pet that says:

"Maybe today is a good day to rest."

Or:

"That's okay! We can try again tomorrow."

It should celebrate achievements without punishing people for difficult days.

And if it could remember previous conversations and progress over time, it might start to feel like a companion I could keep coming back to.

Of course, today's hackathon is short, so building all these features might be unrealistic.

But if I could at least set a goal, report progress, and see the pet react, that would already be fun.

That's what I was imagining this morning.

But Where Would the Sponsor Technologies Actually Fit?

At today's hackathon, participants are encouraged to build with sponsor technologies.

That got me thinking about ai& Inference and Neo4j.

And this is where I got stuck.

I could imagine my AI pet.

I had a rough idea of what the screen might look like.

I could even picture its expressions and animations.

But where exactly would these two technologies fit into the app I wanted to build?

I couldn't immediately connect the technical tools to my idea.

So I asked ChatGPT.

AI Thinks, the Database Remembers, and Code Makes Things Move

After hearing the explanation, things started to make more sense.

Imagine someone telling their pet:

"I published a blog post today!"

An AI model could understand that message and generate a natural reply.

A database could store the user's goal, confirmed progress, and previous events.

And ordinary application code could calculate the completion percentage, update the screen, and make the pet jump.

For this particular idea, ai& Inference could provide access to the AI model, while Neo4j could store goals, progress, and relationships between pieces of information.

In other words, AI doesn't have to do everything.

Maybe that sounds obvious to someone with a technical background.

But I hadn't really connected these pieces before.

Using AI to help me build an app and designing an app that actually uses AI as one of its features are two different things.

They sound similar, but they're not quite the same.

Would the App Stop Working After the Hackathon?

There was another question on my mind.

If I used the sponsor technologies to build something today, what would happen after the hackathon?

Would the app stop working if I could no longer use those services?

I asked ChatGPT about that too.

And I learned that the code we create and the external AI services an app connects to are separate things.

Of course, depending on pricing and usage limits, an AI-powered feature might stop working if access to a service ends.

But if the app is designed so that the AI provider can be changed, it may be possible to keep using the app with a different provider.

That doesn't mean the switch would always be effortless. The API connection and other parts of the code might need changes, and the replacement service might have its own costs.

Still, that was an interesting discovery.

An app built during a hackathon doesn't necessarily have to become a one-day project.

Maybe What I Need to Learn Is How to Choose the Right Technology

Thinking about all this, I realized something about myself.

I've gained quite a bit of experience building apps with AI.

I enjoy coming up with ideas, designing screens, and making characters move.

But I'm less confident when it comes to deciding how AI models, APIs, databases, and ordinary code should work together.

Which part really needs AI?

Which part should use a database?

What can be handled by regular code?

Until now, I've often been able to ask Claude Code or Codex to build something and get a working result.

But going forward, I want to understand why a particular technology is being used, so I can make better decisions myself.

That's a little different from wanting to write every line of code on my own.

I want to understand how the things I imagine can actually be built.

So, What Will I Build Today?

Anyway, it's still the morning before the hackathon.

I haven't built the AI pet.

I don't even know what team I'll end up working with or what we'll eventually decide to create.

But I've already discovered something I want to learn today.

Not just how to build things with AI, but how AI actually fits into the things we build.

That's something I'd like to pay attention to during the hackathon.

Now I'd better finish getting ready!

I'm excited to see what today brings!