Reimagining Ocean Data Workflows with Agentic AI
Date Published
Blue Lobster has been exploring a simple question with profound implications for marine science. What if accessing and analysing complex ocean data felt less like software engineering and more like scientific thinking?
In collaboration with SOCIB (The Balearic Islands Coastal Observing and Forecasting System), the team has developed an AI prototype designed to reduce the friction that has traditionally accompanied ocean data workflows.
From Query to Insight
It begins with what appears to be a straightforward request. Seawater temperature from fixed platforms in July 2024…
Behind the scenes, however, the task is anything but simple. The system must interpret the scientific meaning of the request; identify the correct parameter; determine which observing platforms hold the relevant measurements; and query APIs across multiple services. It must also handle a wide range of marine data formats, including WMS, WMTS, NetCDF, Zarr and OpenDAP.
Only once this orchestration is complete does the user see the outcome: a clean visualisation, downloadable datasets, and a reusable asset stored within the session.
Not long ago, this would have required manual catalogue searches, API scripting, data wrangling in Python or R, and visualisation libraries such as Matplotlib inside Jupyter notebooks. It was specialist work; often time-consuming; occasionally frustrating. The prototype compresses that workflow into a natural, conversational interaction without sacrificing scientific control.
The service is aimed at multiple user types: an early-career researcher exploring a dataset for the first time does not approach a problem in the same way as a senior ocean modeller preparing material for publication. The system recognises this through distinct user personas. In Explorer mode, the emphasis is on guided discovery and clarity; parameters are surfaced helpfully, and outputs are framed to support understanding. In Analyst mode, the tone shifts towards precision and efficiency; assumptions are minimised, processing levels are made explicit, and comparable datasets are prioritised.
Comparison Without Complexity
The real power becomes evident when questions evolve.
How does July 2024 compare with July 2025?
The system infers that the same stations should be used, identifies comparable processing levels, retrieves matching datasets, and presents a structured comparison. The scientist can explore the visualisations directly or ask the system to summarise the differences observed.
In scientific environments, predictability matters; AI cannot be speculative. Outputs must be consistent, traceable, repeatable, and grounded in validated datasets. The emphasis here is on reliability and transparency rather than novelty.
Seeing the Ocean in Context
Marine science rarely exists in a single dimension. Location shapes interpretation.
The prototype allows spatial reasoning to sit naturally within the workflow. A buoy’s position can be mapped instantly. Salinity data west of Mallorca can be retrieved through natural language. Satellite layers from Copernicus can be added for a broader environmental context.
Biodiversity data can also be woven into the same session. Species information from FishBase and distribution records from the Global Biodiversity Information Facility are integrated alongside observational data. A query about red mullet moves seamlessly from culinary curiosity to structured biological insight, complete with imagery and mapped distribution.
Depth, Movement and Multi-Dimensional Data
Ocean data is rarely static.
When switching from fixed platforms to gliders, the system adapts. Salinity measurements are visualised across depth as the glider travels through the water column. Temporal, spatial and vertical dimensions combine into a coherent view that would previously have required substantial manual preparation.
The prototype does not remove scientific judgment. It removes the mechanical burden that can stand between question and exploration.
Knowledge at Hand
The system also draws upon document-based knowledge stores. Questions about quality control procedures for glider data, for example, return structured summaries grounded in source documentation. Citations are surfaced; follow-up actions are suggested. The workflow blends data retrieval with methodological context.
The knowledge base is not limited to internal notes or informal guidance. It can incorporate technical manuals, platform documentation, metadata standards, operational protocols and peer-reviewed publications relevant to the observing system. Scientific papers describing calibration methods, validation studies, sensor limitations or processing pipelines can be indexed alongside official documentation.
This allows the AI to do more than summarise text. It can connect a dataset to the methodological framework that underpins it; explain how quality-control flags are applied; clarify what a specific processing level implies; or reference the original publication that defines a measurement standard.
By grounding responses in authoritative sources, the system supports interpretability and traceability. The aim is not to replace scientific literature, but to surface it as it becomes relevant, helping researchers move fluidly between raw observations, processing context and the broader body of scientific evidence.
Where This Could Lead
The current focus is on developing reliable tools and workflows for diverse marine data types. Yet the trajectory is clear.
Trend detection, anomaly identification, integration with predictive models, and scenario testing. These are natural extensions of the same agent-based foundation.
Watch the Full Demonstration
To see the prototype in action and explore the workflows described above in greater detail, watch the full presentation on YouTube.
The demonstration walks through real-world scientific queries, platform comparisons, biodiversity integration and spatial analysis, offering a clear view of how agent-based AI can transform marine data workflows.
Watch the full presentation here: https://youtu.be/NLYehsAPPDc