How a Google Gemini Hackathon Convinced Me to Build My First Chrome Extension.

SeaSussed was submitted to the Gemini Live Agent Challenge hackathon hosted on Devpost — built by someone with no computer science background, using Claude Code. Here’s why a willingness to be confused can go a long way and why you shouldn’t be afraid to try it yourself.

Let Me Tell You Who I’m Not

I am not a software engineer. I’m not a computer science graduate. I don’t have years of coding experience or a GitHub history full of impressive repositories.

I’m an expat from the United States, currently living in Asturias, Spain. My background is in clinical care, public health, and pharmaceutical studies. I’ve had a few stints working with startups, but most of my growth over the years is actually from being an immigrant. I have learned to be resilient, to handle unknowns, to adapt to new situations, and (most importantly) to be uncomfortable. To be quite honest, going outside my comfort zone wasn’t always a choice; however, it’s been the key to my personal growth and confidence.

I’m writing this because I think there are many people who are thinking “I’d love to build something, but I don’t know where to start” or “I’m not technical enough for that.” And I think it matters that there are many people who thought the same and decided to do it anyway. And that AI skills and tools shouldn’t be something controlled by few.

Hackathons are an amazing way to learn and practice these tools. So far, I’ve participated in two Google Gemini hackathons this year. This is why these are worth considering (especially if you’re new) and why I will participate with them again:

They give you the tools for free. The hackathon provides quickstarts, tutorials, access to the Agent Development Kit, and expert-hosted webinars Google Cloud — so you’re not dropped in cold. You have a genuine learning path built into the experience.

It’s a level playing field. The challenge is open to everyone, whether you are a seasoned AI engineer or writing your first line of code. Gemini 3 Hackathon That’s not marketing language (I’m living proof of it).

Early API access. Participants get hands-on access to new Gemini APIs before the rest of the world catches up. Gemini 3 Hackathon For someone learning, this is genuinely exciting because you’re building with cutting-edge tools at the same time as professionals who do this for a living.

The prizes are real. The grand prize includes a trip to Google Cloud Next ’26 in Las Vegas, $25,000 in USD, $3,000 in Google Cloud credits, and the opportunity to present on stage. Google Cloud But honestly, the less glamorous prizes mattered more to me — the Google Cloud credits let you actually deploy something without a bill.

You get community. Entering connects you to Google Developer Groups and a global community of builders at every skill level. That network has value long after the deadline passes.

So there is the background on why I chose this hackathon and what gave me the courage to enter. Then came the idea… that came from a visit to the Monterey Bay Aquarium, a SeaWatch card, and a desire to do better.

The Story of SeaSussed

Years ago, my older sister took me to the Monterey Bay Aquarium. I was completely blown away by the jellyfish, the kelp forests, the quiet intelligence of the sea otters. It’s one of the best aquariums in the world. It’s also a unique opportunity to educate and to promote activism.

When we left, the aquarium staff handed us a Seafood Watch card: a small printed guide to help newly inspired visitors make better consumer choices. It was a genuinely good idea. But it was also just a tiny paper scorecard. Static, vague, and unable to be applied to specific cases. It’s designed that way and it worked for a broad application of some general recommendations in an effort to reach as many people as possible. The issue is, we know not all products are equal and static information didn’t result in confident decisions.

What I actually needed was specific, up-to-date information about the exact product in front of me. I also needed more education. There are many claims and certifications thrown around in the food industry today, but I didn’t know what any of these labels meant, nor the impact.

SeaSussed is, in many ways, my answer to that card.

SeaSussed is a Chrome extension that gives you an instant sustainability score for any seafood product while you’re shopping online. It leverages AI to pull in all the important factors, to research, to assess, to translate information into an understandable summary, and to answer questions about the very product you’re interested in — not “salmon” in general, but that filet of wild-caught Alaskan salmon that went on sale for $15.99 a pound. Is this a “good” choice? Well, at least now I have the information I need to feel more confident in that decision.

How I Actually Built This (The Honest Answer)

I built SeaSussed a little over the course of about a month, using Claude Code as my primary development tool. I described what I wanted in plain English. I asked questions when I didn’t understand what was happening. I made decisions about the product (what it should do, how it should feel, what mattered most) and Claude helped me turn those decisions into working code.

I want to be clear about what that process actually looked like, because I think people imagine it’s either effortless (“the AI just does it for you”), somehow dishonest (“you didn’t really build it then”) or still out of their league (“I’m not technical enough for that”).

Building with AI assistance still requires you to understand the problem deeply. It requires you to make judgment calls at every fork in the road. It requires you to test, notice when something is wrong, articulate why it’s wrong, and keep iterating. The AI doesn’t do any of that for you. However, it does lower the barrier between your idea and working code.

There were moments I felt completely lost. There were architecture decisions I didn’t fully understand until I’d broken them twice. There were features I had to cut because I couldn’t get them working in time. SeaSussed is not a perfect product. But it is a real one. It’s submitted, functional, and doing something genuinely useful. And it was built by your average mom of two toddlers who, a few weeks ago had never written a Chrome extension.

What SeaSussed Does Under the Hood

SeaSussed is a Chrome Manifest V3 extension paired with a Python FastAPI backend deployed on Google Cloud Run.

When you land on a seafood product page, the extension captures three signals simultaneously: a viewport screenshot, product gallery images, and the page’s DOM text. All of it gets sent to the backend in a single request. This multi-signal approach was deliberate in order to capture all of the back-of-package certifications, origin labels, and secondary images that a single shot would miss, and often carry the most useful sustainability information.

The backend runs Gemini 2.5 Flash on Vertex AI to extract structured product information from that combined visual and textual data. It then scores the product using a Python scoring engine backed by a SQLite database built from FishBase, NOAA FishWatch, and MSC certification data.

Results stream back progressively via Server-Sent Events. The score appears while alternatives and explanations continue loading in the background. With tools like these, it’s important to keep the experience feeling fast even when the full analysis takes a few extra seconds.

The Voice Layer: Talking to Your Browser

Voice mode is where SeaSussed gets genuinely interesting.

Using Gemini Live (gemini-live-2.5-flash-native-audio) over a WebSocket connection, users can speak to SeaSussed directly. The voice agent has three tools: analyze the current product page, search the store for alternatives, and navigate the browser to a product. You can say “Is this salmon sustainable?” or “Find me an option with a better score” and it acts.

This was the part of the build that humbled me the most. Getting a voice agent to behave reliably requires a kind of iterative debugging that takes a lot of patience and critical thinking. You have to think about when the tool should be called, navigation reasoning, rules to put into place and how to handle timeouts. I’m far from an expert, but it’s from the challenges that you learn the most.

Three Problems That Made Me a Better Builder

1. Multi-image vision

The first version only sent a single screenshot. It was obvious it wasn’t enough because it failed to bring in all of information from all sides of the package in secondary gallery images. The fix was to scrape the product gallery URLs, fetch each image as base64, and pass everything to Gemini alongside the DOM text. This was one of the first moments I realized that working with AI models isn’t just about calling an API. You have plenty of opportunities to think carefully about what information you give it and how you frame the question.

2. Scoring from listing text alone

When the voice agent searches for alternatives, it only has listing text. There are no images, no certification badges. This produces lower score estimates than a full-page analysis. Early versions used these estimates as a threshold and blocked almost everything. The fix was to remove the threshold and treat the ranking as directional. With the perspective of expectation versus reality, I would say this is where the first version has fallen short of my aspirations. However, I haven’t given up on improvements and this is an area where I will continue working.

3. Progressive streaming

Making the UI feel fast required splitting the backend into two phases: scoring first, then alternatives and explanation. Coordinating these phases with the frontend animation was one of the most satisfying improvements because from a user perspective, second feels long, and a second with an unknown tool might as well be an eternity. You can easily lose people before they even try what you built because of a long latency. Putting yourself in the shoes of the user and thinking of ways to make these wait times less painful is exactly the kind of creativity and human necessity that is needed in AI coding today.

To Anyone Who Is Nervous to Start

If you’re reading this article and some of the terminology is unfamiliar, I just want you to know that I was in your shoes not too long ago. Three months ago, I would have read some of these paragraphs and would have seen the words but would have lost the meaning. Learning to use the AI tools available today (many even offer a free tier), and then building something that uses them, is exactly the kind of journey I encourage you to take yourself.

Here is what I want you to know: You need to understand your problem well, care enough to keep going when things break, be willing to ask questions and find answers. Discomfort and struggle is part of learning.

SeaSussed is far from perfect. There are features I cut. There are things I would rebuild differently now. There is latency I haven’t solved. But I entered a Google AI hackathon and I submitted something real.

You should do it too.

The Stack

Frontend: Chrome Manifest V3 extension (JavaScript, side panel API)

Backend: Python FastAPI, deployed on Google Cloud Run

AI: Gemini 2.5 Flash via Vertex AI (vision + extraction), Gemini Live for voice

Development tool: Claude Code

Data: SQLite from FishBase, NOAA FishWatch, and MSC certification records

Why This Matters Beyond the Hackathon

Overfishing is one of the most under discussed environmental crises happening right now. Certification systems exist, but they’re fragmented and invisible at the point of purchase. SeaSussed is a small attempt to close that gap and to make the sustainable choice as easy as the default choice.

But beyond the environmental mission, SeaSussed represents something I care about equally: the democratization of building and creation. Having a good problem to solve and the drive to solve it can get you really far.

You can view the full project at devpost.com/software/seasussed. And if you’re working on something… however unfinished, however uncertain… I’d love to hear about it.

— — —

#GeminiLiveAgentChallenge #GirlsWhoCode #WomenInSTEM