What if one platform could verify them all?

That question stayed in my head longer than I’d like to admit.

So instead of building another CRUD application with a login page and a dashboard (the internet has enough of those), I decided to build PARAKH—an AI-powered verification platform designed to detect scams across multiple mediums: voices, websites, documents, QR codes, emails, and images.

Spoiler: My RAM deserved financial compensation.

⸻

The moment everything clicked

Most projects begin with an assignment.

This one began with paranoia.

Imagine getting a call from your mother.

Same voice.

Same tone.

Same way of saying your name.

She asks you to transfer money immediately.

You wouldn’t think twice.

Until you realise…

It wasn’t her.

With today’s AI voice cloning technology, that’s no longer a far-fetched scenario.

Instead of asking,

“Can I detect AI?”

I started asking,

“Can I verify trust?”

That single question changed the entire direction of the project.

⸻

Suddenly, one model wasn’t enough.

Voice verification quickly turned into document verification.

Document verification turned into website analysis.

Website analysis turned into QR verification.

Then came email authentication.

Image forensics.

Metadata analysis.

An LLM that could explain every result without sounding like a research paper.

Somewhere in between, Chrome had 47 tabs open and my laptop sounded like it was preparing for takeoff.

⸻

The biggest lesson wasn’t technical.

Training models is exciting.

Connecting APIs is satisfying.

Deploying successfully feels illegal.

But none of those matter if users don’t understand the result.

Imagine seeing:

Confidence: 93.81%

Cool.

Now what?

Instead, I wanted PARAKH to explain things like:

“This website was registered recently and contains characteristics commonly associated with phishing.”

or

“The uploaded voice does not closely match the enrolled speaker.”

Technology shouldn’t just be intelligent.

It should be understandable.

⸻

Debugging has a strange sense of humour.

Every developer has that one bug.

Mine arrived five minutes before a demo.

Everything had worked perfectly.

Until…

It didn’t.

After three hours of debugging, checking logs, questioning my career choices, and blaming perfectly innocent APIs…

The culprit?

One mismatched field name.

That’s it.

Three hours.

One missing word.

Character development completed.

⸻

Things college never really teaches you

Nobody tells you that software engineering is mostly reading documentation.

Or debugging.

Or redesigning something you were absolutely convinced was “perfect” yesterday.

Or learning an entirely new technology because your project suddenly needs it.

Or discovering that the bug was introduced by… yourself.

Past You is a dangerous developer.

⸻

My favourite feature wasn’t the AI.

It was the question behind it.

Instead of asking,

“Is this audio AI-generated?”

I asked,

“Is this actually the person they claim to be?”

That tiny change completely reshaped the architecture.

Sometimes better software doesn’t come from better models.

It comes from asking better questions.

⸻

If I rebuilt PARAKH today…

I’d spend less time chasing perfect code.

And more time designing systems that are easier to extend, maintain, and debug.

Because clever code impresses developers.

Clean architecture survives developers.

⸻

Final thoughts

Technology changes ridiculously fast.

Today’s impossible idea becomes tomorrow’s tutorial.

There’s always another framework.

Another model.

Another API.

But curiosity never goes out of date.

So I’ll keep building.

Keep learning.

Keep shipping.

And hopefully spend slightly less time hunting missing field names.

Although, realistically…

Probably not.

⸻

If you’ve made it this far…

Thank you.

If you’re building something ambitious too, don’t wait until you feel “ready.”

You probably won’t.

Start anyway.

The bugs are already waiting for you.

⸻

P.S.

The best part about building AI isn’t teaching machines to think.

It’s learning how to ask better questions yourself.