Project Overview
A lightweight, deterministic inference runtime designed specifically for space-rated flight computers, enabling complex AI models to run in the harsh radiation of space.
The Problem
Modern AI frameworks like PyTorch and TensorFlow are too bloated and non-deterministic to run on the highly constrained, radiation-hardened (rad-hard) processors used in modern satellites.
Our Solution
We developed a custom, bare-metal C++ inference engine that compiles trained neural networks into highly optimized, deterministic static binaries. It supports quantization, pruning, and memory-safe execution with zero dynamic allocation.
System Architecture
The framework takes ONNX models from standard training pipelines and compiles them ahead-of-time (AOT) for specific Rad-Hard architectures like the LEON4 or ARM Cortex-R.
Impact & Results
Achieved 40x faster inference and 90x lower memory footprint compared to standard edge runtimes. Successfully deployed on a flagship lunar mission for real-time crater detection during descent.