I Built My Personal Portfolio with Gemini 3.8 Flash High In A Weekend

What happened when I stopped coding every component manually and let an AI coding agent take the first shot.

There was a point where I realized something slightly embarrassing.

I’ve been building software for years. I’ve worked with JavaScript, TypeScript, React, Next.js, Node.js, Java, Python, micro-frontends, cloud platforms, and lately, AI agents. I’ve also written about technology for years.

And yet my own portfolio was still something I kept postponing. Not because building a portfolio is technically difficult, because every developer knows the trap: “I’ll make it when I have time.”

Then you never have time.

So last weekend Sunday Afternoon, I decided to change the approach.

Instead of spending days designing every component and writing every line manually, I wanted to see how far I could get using Antigravity with Gemini 3.8 Flash High.

The goal wasn’t to create another AI-generated template. I wanted to see if I could build something that actually felt like me. And surprisingly, the result was much better than I expected.

I wanted a personal portfolio that communicated three things quickly:

I’m a full-stack engineer.

I build modern web architectures.

I’m increasingly focused on AI, agents, and emerging developer technologies.

But I didn’t want the typical developer portfolio.

You know the one has

I wanted something much simpler. The final structure became:

That’s it.

Before asking Gemini to build anything, I already had a visual direction.

I wanted something:

The hero was particularly important.

Instead of using a generic developer illustration, I wanted an actual representation of myself working at a developer workstation.

The final hero uses a dark cinematic environment with my portrait integrated into a workstation scene. I also wanted the hero to occupy the entire viewport. That small change made a surprisingly big difference.

The first screen now feels more like an introduction than a website header. And I added a subtle scroll indicator in the middle-bottom area to make it obvious that there’s more below.

This was where the experiment became interesting. I gave Gemini the overall requirements rather than individually asking it to build:

Instead, I gave it the entire product vision. The stack was straightforward: Next.js + Tailwind CSS

The prompt described:

Then I let Gemini take the first implementation. And this is where AI coding agents are becoming genuinely interesting.

And that’s important. It shouldn’t be. The first implementation had things I didn’t want.

This is exactly where I think the mindset around AI coding needs to change.

The goal isn’t: “AI generated perfect code.”

The goal is: “AI gave me a strong first implementation that I could review, direct, and refine.”

Those are very different things.

One of the biggest lessons from the experiment was around content accuracy.

For a portfolio, hallucinated content is especially dangerous.

So I manually provided the real links and content that mattered.

The lesson was simple: Let AI generate implementation. Don’t let it invent your identity.

Another interesting iteration was the technology section. The first version used placeholder-style icons, that wasn’t good enough.

If I’m listing JavaScript, TypeScript, React, Next.js, Docker, PostgreSQL, GitHub, and other technologies, I want the actual technology/brand icons.

So I explicitly directed Gemini to use two specific sources Devicon for developer technologies and React Simple Icons for broader brand icons.

That distinction matters. A real technology logo is much better than an AI-generated approximation.

One of the biggest surprises came after deployment is I checked the website on desktop, Looks good.

Then mobile, Looks good. Then tablet, It was actually better than I expected.

Tablet is often the awkward breakpoint, Desktop layouts shrink and Mobile layouts stretch.

Tablet sits somewhere in between and exposes bad assumptions. But the final site translated surprisingly well across all three.

That was one of the moments where I realized:

It can also help produce a reasonable first-pass responsive system when the requirements are clearly communicated.

Of course, that doesn’t eliminate testing. It makes testing more valuable.

I didn’t want to simply build the site and immediately publish it. Before deploying, I wanted to check for:

This distinction is important.

A public portfolio can intentionally contain:

That’s not necessarily a security problem. But accidentally publishing:

is a very different situation.

AI-generated applications still need normal engineering discipline.

The final application is now live on Vercel: vijaydeepak.vercel.app

And that’s where the experiment became real.

Not that“Gemini generated some code”, but “I described a product, iterated on it, validated it, and shipped it.”

Here’s another interesting data point.

I used Gemini 3.8 Flash High for the full development workflow in Antigravity and after completing the portfolio, my usage dashboard showed:

Weekly limit remaining: 90%

Five-hour limit remaining: 37%

I don’t want to turn that into a scientific benchmark. It isn’t one.

Different prompts, context sizes, models, agent behavior, and projects will produce very different results.

But as a personal experiment, it was interesting.

I was able to go from: nothing → designed portfolio → responsive implementation → iterative fixes → security review → production deployment

in a weekend without burning through my weekly allowance.

The biggest lesson wasn’t that Gemini can build websites.

We already know that the interesting lesson was about how you use it.

A vague prompt produces a generic website. A detailed prompt containing:

produces something much closer to what you actually want.

Gemini can decide how to implement a card.

It shouldn’t decide:

Those are source-of-truth problems.

The workflow wasn’t: Prompt → perfect website

It was: Prompt → implementation → inspect → correct → refine → validate → deploy

That’s much closer to real engineering.

Once the portfolio looked right, I stopped asking the AI to “make it better.” That’s a surprisingly important discipline. AI agents are very good at changing things. That doesn’t mean everything needs changing.

Absolutely, but I wouldn’t use the exact same workflow for every project.

For a marketing site, portfolio, landing page, prototype, or content-driven application, this approach can be extremely productive.

For a production system involving:

I’d use AI very differently.

I’d want much tighter architectural control, testing, threat modeling, code review, and human validation.

The more consequential the software becomes, the less “just let the agent build it” makes sense.

The portfolio itself wasn’t really the experiment. The experiment was testing a different development workflow.

Instead of spending the weekend writing every component manually, I focused more on:

What should exist?

Why should it exist?

What should it look like?

What must not change?

What information is authoritative?

Then I let the AI handle much of the implementation. That feels like an important shift for software engineers.

We’re gradually moving from writing every line toward directing, reviewing, validating, and shipping systems.

But that doesn’t make engineering less important. It makes engineering judgment more important.

I started the sunday afternoon with a portfolio I kept postponing. I finished with a live website.

And the most interesting part wasn’t the speed. It was how much of the work moved from typing code to making decisions.

That’s probably where AI-assisted development gets really interesting is not when AI writes 10,000 lines of code.

But when a developer can describe an idea clearly, guide an agent through implementation, catch its mistakes, preserve the important constraints, and still take responsibility for what gets shipped.

AI didn’t build my portfolio for me.

I directed it, reviewed it, corrected it, and shipped it. And for a weekend experiment, that’s a pretty good result.

I hope you found it helpful and informative. If you have any questions or feedback, feel free to leave a comment below. Your support and engagement mean a lot to me.