OnAIR GPU Plugin

GPU-accelerated inference plugin for NASA OnAIR cognitive framework.

This plugin enables OnAIR to consume results from GPU-based geospatial foundation model inference running on NVIDIA Jetson hardware. It implements the Complex Reasoner interface, correlating inference outputs with satellite telemetry to produce prioritized action queues.

Architecture

The plugin follows OnAIR's external plugin model:

• Inference pipeline processes imagery on the GPU independently
• Results (classifications, confidence scores, coordinates) are published to Redis
• OnAIR's RedisAdapter ingests results as telemetry data products
• This plugin (Complex Reasoner) correlates results with operational context and triages actions

No images flow through OnAIR's data pipeline. No modifications to OnAIR's codebase are required.

Status

Design phase. See docs/DESIGN.md for the full design document, currently under review.

Integration Pattern

GPU Inference Pipeline  ──► Redis ──► OnAIR RedisAdapter ──► GPU Plugin (CR)
(PyTorch/TensorRT)          │         (existing)              (this repo)
                            │
                      JSON messages:
                      - detection_class
                      - confidence
                      - coordinates
                      - severity
                      - gpu_health

Requirements

• NASA OnAIR (tested with latest main)
• Python 3.9+
• Redis server
• GPU inference pipeline publishing to Redis (e.g., GodelEDGE)

Quick Start

# Clone this repo into your OnAIR plugins directory
git clone https://github.com/godel-space/onair-gpu-plugin.git plugins/gpu_inference_reasoner

# Configure OnAIR to use the plugin (see docs/DESIGN.md Section 5)
# Point ComplexPluginDict to the plugin path in your config.ini

References

• NASA OnAIR
• OnAIR Architecture
• IAAI25 Paper
• ELISA Aerospace WG GodelEDGE Profile

License

Apache License 2.0

Author

Godel Space -- Autonomous AI agents for satellite Earth observation.