At Vietnam Japan AI Hackathon 2026, our two-person team, FrostLoop AI, made it into the Top 16!
Of course, I was really happy.
But when I read the judges' feedback afterward, there were a few comments that made me stop and think.
"How is your product different from Amazon's existing services?"
Oh, good point!
Even if we think our idea is interesting, it's hard to explain what makes it unique when similar services already exist.
The judges also pointed out that, although we were building an AI agent, our product was focused more on giving advice than actually performing actions.
So I started researching Amazon's existing services and looking back at my own experience as an Amazon seller.
It was time to rethink what could make FrostLoop AI different.
This is the story of how challenging it can be to develop an AI product at a hackathon — and how my everyday business experience turned out to be unexpectedly useful.
We Made the Top 16! But There Was Still Plenty to Improve
FrostLoop AI is an AI agent designed to address problems faced by people who sell products on Amazon.
I sell products on Amazon myself, and a problem I'd experienced in my own business became the starting point for this project.
My teammate is responsible for the AI agent development.
Meanwhile, I'm contributing my business experience and perspective as an actual Amazon seller.
We're approaching the same project from two different directions.
In the Top 16 selection feedback, the judges recognized that we were addressing a real-world problem.
They also gave positive comments about the technical architecture of our AI agent.
But they identified areas that needed improvement.
Two points stood out to me.
- We hadn't clearly shown how FrostLoop AI was different from Amazon's existing tools.
- If the AI mainly recommends actions rather than carrying them out, its capabilities as an agent may be too limited.
Those were pretty sharp observations!
What surprised me was that we were receiving this kind of feedback at such an early stage.
And I was actually grateful for it.
Once you know what's missing, you have something concrete to think about and improve.
Amazon Has Similar AI Services
So I started looking into the tools Amazon already provides.
That's when I discovered that Amazon itself has been developing AI services to support sellers.
One example is Seller Assistant.
It's designed to help sellers with tasks such as reviewing inventory and sales information.
And Amazon is moving beyond AI that simply makes recommendations. It's also developing capabilities that can take certain actions with the seller's authorization.
Wait, Amazon is already doing that much?
Even though I sell on Amazon, I don't keep track of every new feature the company introduces.
Researching the competition helped me learn more about what Amazon itself is working on.
Of course, just because Amazon has announced a feature doesn't mean it's available to every seller in Japan.
Still, one thing became clear.
If our product only analyzes inventory and recommends what sellers should do next, explaining what makes it unique could be difficult.
That was a pretty important discovery.
Is Discounting Products to Clear Inventory Always the Best Answer?
While researching competing services, another thought came to mind.
I sell products under a brand that I've developed myself.
So simply reducing inventory isn't always my only priority.
Let's say a product isn't selling well.
Lowering its price might help it sell faster.
But what if there's a price I want to maintain for my brand?
What if I've intentionally limited where my products are sold?
Deciding where to sell a product and how much to charge is also part of running a business.
I have specific sales-channel restrictions as part of my own business strategy.
So if an AI suggested, "Why not sell these products on another marketplace?" that recommendation might not fit my plans at all.
That got me thinking.
The fastest way to reduce inventory isn't necessarily the best choice for every seller.
What if an AI could consider not only inventory levels but also a seller's business strategy and brand priorities?
For example, a seller might set conditions such as how much they're willing to discount, which sales channels they want to use, and which disposal methods they'd prefer to avoid.
Then the AI could compare possible actions within those conditions.
That could be an interesting direction.
Of course, this is still an idea we're exploring.
It doesn't mean all these features have already been implemented in FrostLoop AI.
And it certainly doesn't prove that this approach would be enough to differentiate us from Amazon's existing services.
But I found it interesting that my own experience as a seller gave me another way to think about the problem.
Should an AI Agent Do More Than Just Make Recommendations?
Then there was the second issue raised by the judges.
How much should our AI agent actually be able to do?
FrostLoop AI analyzes data and compares possible actions, such as continuing to sell a product, adjusting its price, or having inventory returned.
But if it only displays recommendations, isn't it closer to an analysis and advisory tool than an AI agent that takes action?
That's when we started discussing API integration.
Amazon has something called the Selling Partner API, or SP-API.
With the appropriate permissions, it allows external software to retrieve certain data and update some types of listing information.
So what if FrostLoop AI could also carry out an action after a seller approved its recommendation?
For example:
The AI analyzes inventory data.
It recommends several possible actions.
The seller reviews the options and chooses what to do.
If the operation is supported, the tool executes it and records the result.
Something along those lines.
Of course, having access to an API doesn't mean every action on Amazon can be automated.
There are limitations depending on the API, permissions, and type of operation.
And I don't want an AI to change prices or dispose of products without the seller's approval.
One idea that's particularly important to me is:
AI recommends. Seller decides.
The AI makes recommendations. The seller makes the final decision.
Then, where technically possible, the tool could carry out the approved action.
That might be a way to combine convenience with the seller's own judgment.
There's still a lot to check before we could implement something like this.
But the judges' comments gave us another direction to explore.
The Perspective of Someone Who Actually Does the Job Matters, Too
In this hackathon, my teammate is leading the technical development.
I'm not a professional programmer.
But there are things I understand because I actually sell products on Amazon.
What data do sellers look at?
Which decisions take time?
What would I want AI to handle, and what would I prefer to decide myself?
The judges' feedback made me think about these questions again.
And something else happened.
Amazon fee data that I'd normally think of as just another business expense suddenly became useful evidence for explaining the problem our AI product was trying to address.
That was a new discovery for me, too.
When people think about AI hackathons, they might imagine that programming skills and technical knowledge matter most.
And of course, they're important.
But understanding who will use an AI product, when they'll use it, and why they'd want it matters just as much.
I'm happy that my own experience can contribute in that way.
Wait, Is Amazon's Own Service the Competition?
It finally hit me.
More than a day later!
When people talk about Amazon's existing services, could they mean those recommendation features I often see in Seller Central?
Is that what they're talking about?
I've never relied on those recommendations to make my decisions. I've always made those decisions myself.
To be clear, I still don't know whether the judges were referring to Amazon's own recommendation features or to a different service offered by a third-party company.
But if they meant Amazon's recommendations, that changes how I think about the comparison.
I've always made decisions based on whether a recommendation fits my own business strategy.
Because I believe that what's optimal from Amazon's perspective isn't necessarily optimal for every seller.
That doesn't mean Amazon's recommendations are wrong.
It's just that different sellers have different priorities.
Some want to protect their brand's pricing.
Others have specific rules about where their products can be sold.
So a recommendation from Amazon might not always be the best choice for my particular business.
Ah, now I see!
Making my own decisions instead of following those recommendations had become so normal to me that I hadn't even thought of them as potential competition.
And suddenly, I felt like I had a better direction for responding to the judges' criticism.
Rather than simply duplicating Amazon's existing features, what if we built AI that respects each seller's circumstances and business strategy when comparing and recommending options?
Maybe that could become one of FrostLoop AI's distinguishing features.
Of course, we'd still need to test whether it really offers something different.
But it's funny how an everyday business decision I'd taken for granted turned out to contain a possible clue for our product.
I'm glad I finally noticed it!
Even After Making the Top 16, There's So Much to Think About!
At first, our focus was simply on using AI to turn our idea into something real.
But after making the Top 16 and reading the judges' feedback, I started thinking about different questions.
If similar services already exist, what should we do differently?
Who really needs our product?
How much responsibility should we give an AI agent?
And what would make someone actually want to use what we're building?
Suddenly, there was so much more to think about! (LOL)
Of course, we don't have all the answers yet.
We still need to explore and validate what could make FrostLoop AI different.
But thanks to the judges' feedback, I had an opportunity to think beyond "What can we build with AI?" and start asking "What would people genuinely want to use?"
That's another experience I wouldn't have had without joining this hackathon.
Pitch Day is coming up on October 24!
We're still preparing as a team, and I'm learning new things right up to the last minute.
I wonder what questions the judges will ask us on the day.
I'm a little nervous, but I'm excited, too!