Technologies
The data infrastructure, AI model suite, and public dashboards designed to take raw monitoring readings and turn them into operational decisions and open releases.
Data is only useful if the infrastructure can reason over it.
Pillar 01 is designed to produce a great deal of data: continuous sensor telemetry, robot survey imagery, lab analyses, satellite layers. The technology stack is what makes that data actionable. By design, it indexes everything spatially and temporally, runs the AI inference that turns sensor signals into early-warning alerts, and exposes the result through public APIs and dashboards anyone can use.
Three obligations shape every technology choice in this pillar.
Robot inference at 5 fps. Sensor anomaly detection at 15-minute resolution. The latency budget is set by the ecological response window, which is days at best for bleaching and minutes for storm-pulse contamination.
Every dataset is designed to be queryable through a public API. Every dashboard layer is keyed to peer-reviewed thresholds. Every model weight is licensed for release to the research community under a permissive license.
Tooling is chosen so a partner site in Year 6 can run the same analysis pipeline on the same data formats without re-implementing the program from scratch. Standards-conformant APIs and well-understood open-source components only.
PostGIS as the spatial backbone.
Every reading, every survey image, and every model prediction is designed to be geolocated. The database is built to handle vector, raster, and time-series data in the same schema, expose it through a standards-conformant public API, and version-control every snapshot the AI models train on.
The database itself is PostGIS on PostgreSQL. The public API is an OGC-API Features endpoint, the open standard for geospatial data published by the Open Geospatial Consortium. Pollution data is structured under the EPA WQX standard, biodiversity records under Darwin Core, and sensor QC under QARTOD. Every model checkpoint is designed to link to a hash-versioned snapshot of its training data, so any published result can be reproduced from its inputs.
Seven models, one shared dataset.
The AI stack is designed to train on the same data the program will publish openly. Vision models will run on the robot itself, doing benthic and species classification during the dive. Prediction and forecasting models will run server-side, fusing aggregated telemetry into early-warning alerts. The biorefinery process control model is the one exception, retained as proprietary because it is tied to the biorefinery and bioremediation formulations described under Pillar 03.
SegFormer-B2 reaches 78.3% mIoU on Indo-Pacific coral datasets (Jammalamadaka et al., 2023). The same architecture will be fine-tuned on Hawaiian benthic substrate, with results published as the training corpus grows.
Open by default.
The default position for every output of the program is open release under a permissive license. Four categories of work are licensed for publication, each under the license that best fits its use. A narrow retained set, scoped to bioremediation and biorefinery compositions, is described under Pillar 03.
Continuous water quality telemetry, benthic survey imagery, lab results, and satellite layers, all designed to be queryable through the public API.
The GIS database API, the model training pipelines, and the data ingestion code that will run program operations.
Fine-tuned weights for benthic segmentation, species classification, and the open environmental components of source attribution.
Peer-reviewed publications with manuscripts and data archives, plus the STL and design files for every 3D-printed product made from biorefinery filament.
Dashboards keyed to WHO standards.
Every threshold on the public dashboard is keyed to the WHO Guidelines for Safe Recreational Water Environments and the WHO Drinking Water Quality Guidelines by design. Federal and state regulatory thresholds are still rendered, but as secondary reference layers. The WHO standard is the design target throughout the program, since it sits at the lower-concentration, more protective end of the available benchmarks for almost every contaminant the program is designed to measure.
The peer-reviewed reasoning behind that choice is laid out under Mission.