Skip to content

Astra: A Fault-Tolerant AI Medical Assistant for Space

NASA HUNCH Mini Rack Project -- Mischa Nelson & Urvi Dhenge

Astra is a fully redundant, offline-capable AI medical assistant system designed for deployment aboard the NASA Gateway space station and future deep-space missions. Built entirely by high school students under the NASA HUNCH (High School Students United with NASA to Create Hardware) program using commercial off-the-shelf hardware, Astra demonstrates that enterprise-grade fault tolerance is achievable with accessible components.

The Problem

Astronauts operating beyond low Earth orbit face a convergence of risks that no existing system adequately addresses:

  • Communication delay. While Gateway's lunar orbit has near-real-time communication (~1.3 seconds), future missions to Mars and beyond will face delays up to 24 minutes one way. Astra is designed for both scenarios -- providing immediate assistance regardless of communication latency. In a medical emergency on a deep-space mission, waiting nearly an hour for a single exchange with a ground-based physician could be fatal.
  • Limited crew medical training. Crew members receive basic medical training, but they are not physicians. Complex diagnoses and treatment decisions require expert guidance that may not be immediately available.
  • No reliable internet in space. Deep-space missions will have severely limited or nonexistent internet connectivity. Any system that depends on cloud services is unusable.
  • Hardware fails. Cosmic radiation, thermal cycling, vibration during launch, and component aging all contribute to hardware failures. A single server running a medical AI is a single point of failure.

The Solution

Astra addresses every one of these challenges:

  1. On-board AI inference. The medical assistant runs locally with zero network latency. Astronauts get immediate responses without waiting for Earth.
  2. Medical-specialized models. The system runs AI models specifically trained on medical literature alongside general-purpose models, providing domain-specific guidance.
  3. Fully offline operation. Every component -- the web interface, the AI models, the database, the embeddings -- runs locally with zero internet dependency.
  4. Triple redundancy with automatic failover. Three independent clusters, each on its own power domain, with automatic failover and self-healing. The system survives losing any two of three clusters and recovers automatically when power is restored.

Key Statistics

Metric Value
Total physical devices 9
Independent K3s clusters 3 (6 Raspberry Pi nodes)
AI inference servers 3 (2x Pi 5 + 1x Jetson Orin Nano)
AI models deployed 7 (including 2 medical-specialized)
Database replication 3-way PostgreSQL streaming, sub-millisecond lag
Failover time Approximately 5 seconds
Cascade failover Tested -- survives loss of any 2 of 3 clusters
Internet dependency Zero -- fully offline capable
Recovery after total power loss Fully automatic within 2 minutes

Team

  • Mischa Nelson -- Lead engineer and system architect.
  • Urvi Dhenge -- Project contributor.

Documentation

This wiki provides comprehensive technical documentation of every aspect of the Astra system.

System Design

Hardware

Kubernetes & Deployment

AI & Medical Models

  • Ollama Engine -- Inference server configuration and API
  • Model Selection -- All 7 models, selection rationale, performance benchmarks
  • Medical AI -- Meditron, MedGemma, and the medical use case

Database & Replication

High Availability

Application & Operations

Engineering

Reference