Ocean Pulse

Inspiration :

We go to UCSD, so the ocean is literally right there and once we started looking at the data, it was hard to just move on. The California Current feeds fisheries, supports wildlife, and sustains millions of people, but nutrient runoff is fueling algal blooms that drain the oxygen out of the water and create dead zones. The worst part? You can't see it happening. By the time it looks bad, it's usually already too late. So we wanted to build something that lets anyone not just scientists identify what's going on beneath the surface before it gets there.

What it does:

Ocean Pulse classifies the health of ocean regions into three states — Healthy, Stressed, or Critical based on real historical data. You type in any city or coastal location, and the dashboard finds your nearest ocean region, shows you its health status, breaks down the factors driving that prediction, and generates AI-powered recommendations specific to that region.

How we built it:

Random Forest Classifier trained on 286,819 samples from the CalCOFI dataset. Seven features, 98% accuracy. Documented through Marimo and Sphinx, deployed on Streamlit, with an LLM plugged in for region-specific policy guidance.

Challenges we ran into:

Identifying how to make our product be useable and also practical for everyone. After talking with a few mentors, we realized that our product was only practical for specialists, so we added another tab or features into our product that helps people who know nothing about chemical stuff just simply inputting their locations would easily able to find out about the ocean water conditions

Accomplishments that we're proud of :

98% accuracy of our model. Discovering that phosphate and nitrate alone can predict oxygen stress. Making 70 years of dense scientific data actually usable and visible by anyone.

What we learned:

The biggest surprise was that nutrient data alone carries enough signal to predict oxygen stress and also the warning signs are encoded in phosphate and nitrate levels long before anything looks visibly wrong. We also learned that 39% of all samples fall into the Critical category, which honestly more than we expected. The data to protect the California Current has existed for decades but it just never had a way to reach the people who needed it.

What's next for Ocean Pulse :

We want to expand beyond the California Current to cover more coastlines globally. We'd love to incorporate real-time sensor data so the model isn't just historical but actually live. Better LLM fine-tuning for even more precise regional recommendations is on the list too. And in terms of long term getting this in front of actual policymakers and conservation organizations, because the insights are only useful if the right people are seeing them