chatty-buoy

An edge-deployed AI crewmate for marine safety and situational awareness.

chatty-buoy is a locally-hosted, multimodal AI assistant designed for NVIDIA Jetson hardware. It acts as an always-on watchstander that monitors NMEA data, transcribes VHF radio traffic, and analyzes visual feeds in real-time.

It keeps a vigilant eye on the horizon so you don't have to—though it may occasionally offer an unsolicited critique of your docking maneuvers.

No cloud. No subscription. No latency. Just 100% local intelligence.

Architecture

For a detailed breakdown of the system capabilities, service orchestration, and Jetson Thor integration, please refer to the Quintessential Architecture.

Setup & Usage

1. Requirements

• NVIDIA Jetson AGX Thor (JetPack 7.0+)
• Docker & NVIDIA Container Runtime
• References: Ultralytics NVIDIA Jetson Guide
• micromamba (Environment: chatty-buoy)
• ONNX Runtime GPU (ARM64/CUDA 13): Hardware-accelerated inference for ONNX models (YOLO, Silero VAD) on Jetson.
• TensorRT-LLM (NVFP4): We utilize Jetson Thor's native NVFP4 precision for Gemma 4, drastically increasing LLM throughput and minimizing memory bandwidth bottlenecks.
• Optional: ngc CLI (Install manually if you need to download new Riva models).

  Install NGC CLI

2. Start the Stack

Step A: Application Infrastructure (Docker) Starts Riva (ASR), Triton (Cortex), and Postgres (Memory).

docker compose up -d

Step B: Voice Synthesis The voice service uses Chatterbox-Turbo (350M) by default for expressive, low-latency speech.

• Primary: Chatterbox-Turbo (runs on GPU 0).
• Fallback: Kokoro-FastAPI (82M) is available in src/voice/Kokoro-FastAPI for ultra-low resource environments.

Step C: The Agent Interact with the system.

micromamba run -n chatty-buoy python3 src/agent_reflex.py

Step D: Watchstander Vision Dashboard The containerized Watchstander automatically observes bounding box events, requests scene summaries from Cosmos-Vision, logs Sentinel telemetry, and provides historical video clip playback. Access the interactive web feed live at: http://localhost:8080

Roadmap

We are currently in Phase 3 (Situational Awareness). See Roadmap for the journey ahead, including NMEA integration, Sonar analysis, and Vision capabilities.

License

MIT