This part begins where the excitement usually starts, the day we finally opened the editor and began building.

For weeks we had discussed ideas, features, and impact.

Now it was time for code.

The plan felt simple:

Capture a face → detect regions → analyze skin → generate guidance.

At that moment, we believed the challenge would be implementation — libraries, models, and optimization.

Instead, the challenge was something much more uncomfortable.

We didn’t actually understand the problem.

The First Lines of Code

We began with face detection. It worked almost immediately. A rectangle appeared around the face on the screen, and for a moment it felt like the project had come alive.

Next we extracted regions — forehead, cheeks, under-eye area.

Again, success.

At this stage DERMA.ai looked functional.We had images. We had regions. We had data. So naturally, we moved to the next step: Train a model to detect acne. And that’s where everything stopped.

A Simple Question That Blocked Us

Before training the model, we needed labels. So we asked ourselves a basic question:

What exactly counts as acne?

Not intuitively.

Not visually.

But computationally.

For a human, acne is obvious.

For a machine, it must be measurable.

And we couldn’t define it.

When Intuition Isn’t Enough

We started looking closely at images.

Some red spots were acne.

Some were irritation.

Some were just lighting.

Some dark areas were pigmentation.

Some were shadows.

Sometimes even we disagreed with each other.

If humans themselves rely on context, experience, and judgment

Then how are we expecting a machine to understand it from a single frame?

That was the moment the project slowed down.

Not because coding was hard.

Because the problem wasn’t clear.

We Were Solving the Wrong Question

We thought we were building a classifier.

But we weren’t.

We were trying to convert a human perception into a mathematical definition.

“Detect acne” is a human instruction.

A machine needs something different:

Measure colour variation

Measure texture irregularity

Measure local contrast

Measure clustered patterns

The problem had to be broken into signals before it could become an algorithm.

Only then did it make sense.

The First Lesson

We learned something unexpected that day:

Before machine learning comes problem definition, healthcare problems are rarely clean. They live in grey areas interpretation, context, probability.

The moment we accepted that, DERMA.ai changed from a coding project into a learning process.

We stopped asking:

How do we train the model?

We started asking:

What should the model observe?

What We Didn’t Know Yet

Once we finally defined measurable features, we built the first working prototype.

It produced results.

For a moment, it felt like success.

Until we tested it somewhere else.

Same person.

Different room.

Completely different output.

And that’s when we discovered our next mistake ,the system wasn’t detecting skin.

It was detecting lighting.

Next Part

In the next blog, I’ll talk about the first real failure of DERMA.ai and how a simple change in environment almost broke the entire idea.

Because building AI doesn’t fail dramatically.

It fails quietly.

This was the very reason behind DERMA.ai

When we first decided to build DERMA.ai the idea sounded straightforward.

“Use computer vision to detect acne, wrinkles and dark circles from a phone camera”

We thought Face detection Already exists. Machine learning models exist, Dermatologist data exists.

So How Hard Could It Be?

Very.

This is the story of initial challenges we faced while building DERMA.ai- and the lessons that changed how we think about Ai in healthcare.

The idea looked simple on paper

The core idea was accessibility.

Skin concerns like acne, pigmentation, dark circles and uneven textures are extremely common. Yet visiting a dermatologist means

• Long waiting time

• High consultation costs

• Limited access in smaller cities

We wanted to build a light weight, computer vision based system that could:

• Detect basic skin features using a webcam/phone camera.

• Combine image analysis with basic lifestyle input

• Provide structured, educational advice.

On paper, it looked elegant

In reality, everything broke in the first week

Challenge 1

People questioned the need!

When we first shared the idea about DERMA.ai, the reaction wasn’t excitement.

It was indifference.

“SKIN that’s not important”

“In our time we didn’t even have face wash”

Skin problems are endured silently.

And because they are not life threatening, they are considered unimportant.

But I learned the hard way that “not life-threatening” does not mean “not impactful.”

It can quietly affect confidence.It can shape how you see yourself in the mirror.It can push you toward experimentation without guidance — and sometimes leave permanent damage behind.

DERMA.ai wasn’t born out of a hackathon sprint or a trend in AI.

It was born out of a mistake.

A personal one.

In this first blog, I wanted to explain why DERMA.ai needed to exist — especially in a culture where prevention is rarely prioritized and skin health is either ignored or trivialized.

In the next part, I’ll talk about what happened when we actually tried to build it, the technical failures, the unexpected bias in data, the ethical dilemmas, and the uncomfortable truths about developing AI in healthcare.

Because building AI isn’t just about training models or improving accuracy.

It’s about earning the responsibility to guide people.

And this is only the beginning.