Architecture Decision Record: Off-Grid Local-First AI Autopilot 📶🤖
Status
Decided (May 2026)
Context
Standard AI applications rely heavily on cloud-hosted remote LLM model endpoints (e.g. Gemini, OpenAI). For oceangoing yachts crossing open waters, satellite connection (such as Starlink) can suffer from temporary outages, weather interference, high latency times, or complete network disconnections.
Decision
We decided to architect Athena Helm under a Local-First / Off-Grid Adaptive AI Pattern.
- Client-side Fallback: When networks checkouts fail, the UI instantly defaults to local TypeScript/NMEA pattern checking. Toggling light bars or running standard engine fans continues to work without web latency lag.
- On-Board Centralization: If a local Raspberry Pi 5 core or mobile device is active, we recommend channeling natural language API triggers to a local Ollama server running Qwen2.5-3B or Llama-3-8B. This processes exact voice payloads entirely offline aboard the ship.
Consequences
- Guaranteed operational safety of voice command systems under stormy conditions.
- Absolute privacy: zero crew coordinates or vessel telemetry data leaks to external web hosts.
- Low power consumption options for battery-sensitive sailing passages (Qwen2.5-3B instructs consume minimal amperes on a Pi 5).