aiwf-embedded-edge-ai

Use for embedded systems, microcontrollers, RTOS, SBCs, Jetson, sensors, firmware boundaries, power, thermal, edge inference, and embedded AI deployment planning.

SKILL.md

GitHub

AIWF Embedded Edge AI

Core Rule

Inventory the physical target before choosing architecture. Embedded and edge AI work depends on board, CPU/GPU/NPU, RAM/VRAM, storage, power, thermal envelope, OS/RTOS, firmware boundary, sensor buses, actuator interfaces, and deployment/update path.

Workflow

  1. Identify target class: microcontroller, RTOS device, Raspberry Pi/SBC, Jetson, industrial PC, laptop prototype, or simulation-only target.
  2. Record constraints: voltage/current, battery or mains, thermal limits, enclosure, IO buses, camera/sensor interfaces, motor/control interfaces, storage, boot mode, and update method.
  3. Split responsibilities: firmware, real-time control, Linux services, model inference, logging, UI/API, remote telemetry, and recovery mode.
  4. For edge AI, identify model format, runtime, acceleration stack, quantization, memory budget, latency budget, and fallback behavior.
  5. Validate with cheap probes: version inventory, import checks, build checks, device-tree/config review, dry-run, or small sample inference.

Guardrails

  • Do not assume a microcontroller can run model inference or networking without checking memory, timing, and power.
  • Do not assume Jetson/CUDA/TensorRT availability from a project name. Verify JetPack, drivers, CUDA, cuDNN, TensorRT, and target module.
  • Keep real-time safety/control paths separate from high-latency AI reasoning paths.
  • Require watchdog, safe boot, log retrieval, and rollback/update notes for field devices.
  • Add aiwf-nvidia-cuda-cudnn-sdk for CUDA/cuDNN/TensorRT code details.
  • Add aiwf-robotics-systems for sensor/actuator integration and robot architecture.
  • Add aiwf-networking-iot for telemetry, device networking, MQTT, or customer-site connectivity.

Validation Defaults

Use available local checks only:

nvidia-smi
nvcc --version
python -c "import platform; print(platform.platform())"
python -m pytest <embedded-or-edge-tests>

For microcontroller or RTOS work, prefer build/config validation unless a device is explicitly connected and approved for flashing.

Primary Source Anchors

Verify the exact board, SDK, and firmware toolchain before recommending install, flash, or deployment steps.