How I Built the Perfect Calorie Tracker — and It Changed Almost Nothing

When you want to lose weight, sooner or later you run into calories. Basic energy balance still exists.

But honestly, for a long time I avoided dealing with it at all. It felt like I’d have to weigh food, count everything manually, dig through nutrition databases.

In short — a pain.

The ideal scenario seemed much simpler: there’s a nutritionist, they tell you what to eat, and you just follow the plan.

But I never actually got to a nutritionist... Then AI happened.

And suddenly it became easy. Take a photo of food, get calories. Ask for your daily target, get an answer.

I even built myself a “janky” calorie tracker on top of ChatGPT. And honestly, it would’ve been fine if not for one thing:

It constantly failed at basic arithmetic.

Summing meals together — already a problem.
Calculating something more complex — basically a side quest.

(Which is normal. LLMs aren’t calculators.)

At some point I thought:

I’m a programmer — can’t I just build something that actually counts properly?

So I did.

Now I could photograph food, get accurate calorie estimates, and see the full picture for the day.

And you know what?

It barely changed anything.

I saw the numbers.
I saw I was overeating.
I understood I should stop.

And yet, in the evening, I could still open the fridge and think:

“Eh, it’s fine.”

…and eat something “untracked” anyway.

Just because I felt like it.

That’s the point where everything breaks.

The problem isn’t that AI calculates badly. And it’s not even that it sometimes makes mistakes.

The problem is that it’s not enough.

A calculator, even a perfect one, doesn’t help you make decisions — it only records the fact that you already overeaten.

It does absolutely nothing in the moment when you open the fridge, think “it’s fine,” and are already seconds away from slipping.

That’s where most AI apps fail.

They give you answers.

But they don’t give you a process.

At some point I realized: I didn’t actually need a calculator. I needed the nutritionist I was thinking about from the very beginning.

Not the one who tells you:

“You went over your calories.”

…after the fact.

The one who keeps you on track and intervenes earlier.

That’s when I started looking at the problem differently.

Not:

“How do I calculate calories?”

But:

“How do I help people avoid falling off track?”

That changed everything.

Suddenly it wasn’t just a number anymore — it became a system.

There’s a goal.
There’s a chosen approach (balanced nutrition, high-protein, keto, or whatever style fits you).
And there’s continuous daily feedback adapting around that — sometimes nudging you, sometimes stopping you, sometimes correcting your direction.

Most importantly, it gives you the next step.

Not just:

“You went 400 kcal over today.”

But:

“Okay, today’s already done — here’s what’s smarter to do tomorrow.”

Or:

“You’re drifting outside your target right now — here’s how to avoid spiraling.”

And weirdly enough, that finally became enough to start changing behavior.

Not always.

But no longer “zero change.”

And this is the important part:

It’s not about how smart the AI is.
It’s not about how many models you use.

What matters is something else entirely:

Does it actually help you in the moment you’re making a decision?

Because the difference is simple:

• A calculator shows you what you did.

• A system helps you do something different.

• At that point, it’s no longer “a chat with AI.”

It becomes something that holds the goal, adapts to you, and quietly keeps you from drifting away from it.

Not because there’s “more AI” inside.

But because there’s direction.

And that direction comes from you.

If you’re curious what came out of all this — here it is:

• iOS version: https://apps.apple.com/app/calculorie/id6757242185 :contentReference[oaicite:0]{index=0}

• Android version: https://play.google.com/store/apps/details?id=com.calculorie.app :contentReference[oaicite:1]{index=1}

• Website: https://calculorie.com/ :contentReference[oaicite:2]{index=2}

You can use it for free.

The subscription only kicks in if you really push usage hard — because model inference unfortunately doesn’t pay for itself.