# Oceanio - **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon) - **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST) - **Where:** Online - **Team:** [Aryan Chettri](https://cerebralvalley.ai/u/zenzone) - **GitHub:** https://github.com/aryanchettripkt-eng/Ocean-Embed - **Demo video:** https://youtu.be/4CKfv5u7T8E - **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery - **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5221 Project Description OceanScope AI shows the ocean below the surface. Satellites measure the sea surface every day, but most of the ocean's heat sits 50 to 300 m down, in the thermocline. That layer controls how strong a cyclone gets, how the monsoon behaves and where fish gather in the Bay of Bengal and Arabian Sea, and it is almost invisible. Below the surface, the only direct measurements come from ARGO floats, each profiling about once every ten days. OceanScope AI fills that gap. From surface observations alone (sea surface temperature and salinity, sea level, currents, winds, heat flux and rain) plus nearby ARGO profiles, it reconstructs daily sea temperature at 15 depths from the surface to about 1,000 m, across the North Indian Ocean on a 0.25° grid (about 25 km). The model is an ensemble of three small U-Nets, trained on 1993–2017 of the GLORYS12 ocean reanalysis and fine-tuned on real float measurements. It predicts how each depth differs from normal for that date. On the website you can explore any depth as a 2D map or a 3D cutaway block of the water column with the thermocline traced. Click any point for its depth profile and history. A time machine steps through real cyclones, so you can see the cold wake they leave 50 m down. Since 29 October 2026 it runs live, publishing a new day every morning from near-real-time data. Meta's AI is how people talk to it. "Talk to the ocean" is a guide built on Meta's Muse models. Muse Spark answers questions by calling our data tools and moves the map to show you, and Muse Voice Transcribe lets you ask out loud. Muse Image makes shareable ocean postcards, and SAM 3.1 powers a lasso that outlines any feature you circle. Every number the guide says is checked against our data, so the AI never invents a value. We also built an Android app for Ray-Ban/Oakley Meta glasses, on Meta's Wearables Device Access Toolkit. Stand on a beach, face the sea, tap the glasses and ask "How warm is it out there at 500 metres?". The app uses your location and heading to find the water in front of you, asks our model, and speaks the answer through the glasses, hands-free. We report results honestly, including where we fall short. On the held-out years 2018–2020, the ensemble's error against GLORYS12 is 0.644 °C, 35.1% better than climatology. On a sealed 2021 test run once, our core model (main2_ft) scored 32.4% skill. On a harder test we set before training, against independent ARGO floats in the Bay of Bengal, we tie the GREP reanalysis ensemble but don't yet beat GLORYS12 or ARMOR3D. Every score is on the site's Validation page. ## More from Meta Global AI Developer Hackathon - [ElderLeaf](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5218) - [sanskar](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5219) - [Cheese Lord](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5220) - [Team Crater2Impactor](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5222) - [team_id=1114754401265198](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5223) - [Team Techno](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5224) --- Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5221. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt