What problem does it solve? Choosing and deploying the correct on-device LLM/VLM runtime for D-Robotics RDK boards is error-prone because each board family requires a different stack, and mismatches (like building hobot_llamacpp on an S600) waste hours of debugging. ## Core Features & Use Cases - Board-aware stack routing: Maps each board to its runtime — hobot_llamacpp (GGUF-BPU) for X5/S100, oellm_runtime SDK (.hbm, nash-p) for S600, and legacy hobot_llm (Bloom 1.4B) for X3 4GB. - Full voice pipeline: Builds the ASR→LLM→TTS loop with sensevoice_ros2 and hobot_tts, or the turnkey xiaozhi (小智) assistant. - Environment probing and benchmarks: Ships llm_env_check.sh to detect board, memory, and installed LLM packages, plus llm_benchmark.py for exact TTFT/TPS/memory figures. - Use Case: A developer asks how to run a vision-language chatbot on an RDK X5; the skill recommends an InternVL3 GGUF-BPU model, explains the required ViT encoder plus language GGUF files, and gives the exact ros2 launch command. ## Quick Start Ask the assistant which LLM or voice assistant stack your RDK board supports and how to build and run it.
