NeonSnap

...Then Google Started Throwing 429 Errors at Me

A few weeks ago I launched NeonSnap, an AI photo transformation app where people can change their hair, outfits, backgrounds, even generate “future baby” predictions. Yeah, I know, not the most original of ideas. But I noticed that a lot of people were doing it badly so I wanted to do a premium version that uses Nano Banana Pro.

It actually worked.

People showed up.
They uploaded photos.
They used all their free transformations.
Some even signed up.

And then… 429 errors.

For hours at a time, the API I depended on would hit capacity limits. Not my bug. Not my infra. Just upstream limits.

There’s something uniquely frustrating about watching real users hit a wall you can’t control.

For a day or two, I emotionally detached from the project. I started mentally drafting my “pivot” before I’d even finished version one.

But here’s what I’ve learned so far:

1. Early traction is chaotic

One day: lots of users.
Next day: almost none.
Then a random spike again.

It’s not linear. It’s not rational. It’s not predictable.

2. Infrastructure maturity matters earlier than you think

Gemini API was “fine for prototyping.”
Vertex AI is the “grown up” version.

I didn’t want to switch yet. But waiting would just compound technical debt and credibility damage. So I moved earlier than planned.

I’m currently part of Google for Start-ups and using their cloud credits to run NeonSnap on Vertex AI. The credits definitely help from a cost perspective. But they don’t magically solve scaling issues. You still hit rate limits. You still see timeouts. You still have to design around capacity constraints.

3. Users tell you what your product really is

I thought NeonSnap would be about subtle glow-ups.

Users typed:

• “Make me look like Madonna in the 80s”

• “Turn me into a supermodel”

• “Make me look like a mermaid”

The “Future Baby” feature? That’s currently the most popular.

The feature I care most about? Not even close.

That tension is interesting.

4. I want a lifestyle business

But right now I’d be happy with a few hundred pounds a month.

Obviously though, it would be nice to have more.

Which means this can’t just be “another AI tool.”
It needs a sharp identity and a growth engine.

Right now I’m doubling down on:

• High-quality identity-preserving transformations

• Clear positioning

• Pinterest for distribution

• A proper paywall soon

If you’re building with AI APIs:

How are you handling rate limits and upstream instability without massive budgets?

And more broadly, how do you decide when to double down vs pivot when early traction is noisy?

Would love to hear from others in the trenches.