Cloud to Air Gap
RAIC deploys as containerized services on Kubernetes, enabling a common application architecture across cloud, on premises, classified, disconnected, and fully isolated environments.
RAIC Platform
Find signals across millions of square kilometers.
Results are arranged in a shared vector space by visual similarity. Dense clusters reveal recurring forms and outliers, so analysts can select meaningful groups rather than review a flat ranked list.
- Color-coded similarity from closest matches to visual outliers
- Polygon and individual selection for rapid triage
- Source imagery, sensor, date, and coordinates available in map context
- Selected detections can move directly into the shared taxonomy
The analyst loop
Iterative discovery from minimal examples.
RAIC keeps the analyst in control while reducing the time between a new question and an actionable answer.
-
01
Cold start from one seed
Begin with a single reference image. No labeled training set and no model development cycle required.
-
02
Search time and area
Query the collection across the selected temporal and spatial extent, using efficient CPU compute.
-
03
Explore and capture
Results group by visual similarity. Triage dense clusters and commit the relevant detections to the taxonomy.
-
04
Reseed for precision
Search again from the curated set. Every pass tightens results and builds mission-specific knowledge.
-
05
Extract and disseminate
Export validated detections, labels, and geospatial metadata to downstream systems.
Iterate
Broad-area search
New question. Same imagery. No retraining.
Use one or more reference images to search across locations, time ranges, and large collections without requiring labels or a pre-trained class for every target.
Seed from Anywhere
Upload an example, draw around an object in source imagery, or search from an established taxonomy category.
AOI, TOI, Sensor
Set spatial and temporal extents, select compatible detectors, and revisit searches from history.
Terrain, Structures, Objects
Search fixed-size visual regions when the mission calls for scene-level patterns rather than discrete objects.
Shared taxonomy
Build a living visual intelligence library.
Organize approved detections into a nested, mission-defined taxonomy shared across the workspace. Every analyst decision improves what the team can search, review, and model next.
- Nest, rename, move, approve, reject, and review categories
- Seed future searches from all images in a trusted category
- Inspect every item in its original map and timeline context
- Export categories and subcategories to CVAT 1.1 or GeoJSON
Analyst-built detectors
Rapidly move from discovery to automation.
Build shallow models directly from approved taxonomy categories. A few labeled examples can become a detector that analysts train, version, run, inspect, and refine without handing the mission to an AI specialist.
- Versioned models tied to selected taxonomy categories
- Confidence controls and category filters in the results view
- GeoJSON export with class, confidence, location, source, and model metadata
Technical foundation
Deploy across cloud, edge, and disconnected environments.
RAIC is built for operational environments and integration into existing geospatial systems, not for moving mission data into a disconnected workflow.
Geospatial APIs and Standards
REST APIs, STAC endpoints, and GeoJSON connect RAIC to existing imagery catalogs, geospatial applications, and downstream exploitation workflows.
MCP and Agent Workflows
An MCP server exposes search, taxonomy, and detector operations as callable tools, so AI agents can run imagery analysis and return grounded results alongside analysts.
Event-Driven Data Fabrics
Integrate with messaging systems such as Apache Kafka to consume events and publish detections, alerts, metadata, and analytic outputs into existing data flows.
Technical paper
A generalizable embedding space for geospatial vision.
RAIC is pretrained unsupervised on a broad corpus of remote sensing imagery, including multispectral sources. The result is a compact embedding space that captures spatial and spectral structure once, then supports search, clustering, and classification without a new deep model for every class.
On PatternNet and EuroSAT, a logistic regression head trained on RAIC embeddings surpassed published CNN accuracies using a few percent of each labeled set. Those classifiers trained in 0.2 seconds on CPU. t-SNE plots of the embeddings show the classes are already separable, which is why shallow models and analyst-in-the-loop retrieval can replace GPU-bound retraining for many geospatial tasks.
Download whitepaper
Technical documentation
Open Documentation
Review the platform architecture.
Read analyst workflows, platform concepts, deployment guidance, and API references in the complete documentation.