An automated, multi-agent AI workflow for compiling, deploying, and tuning computer vision models (YOLO, ViT) on Edge devices (RDK X5, ESP32-S3, STM32).
Driven by the Xiaomi MiMo LLM API, this system utilizes a multi-agent architecture (Architect, Deployer, Evaluator) to achieve a complete Hardware-in-the-Loop (HIL) closed loop: generating deployment strategies, cross-compiling, executing via SSH, analyzing hardware logs, and self-correcting errors.
- Architect Agent: Analyzes hardware specs and decides on model architecture and quantization (INT8/FP16).
- Deployer Agent: Generates deployment scripts, handles ONNX exports, and flashes to edge devices via CLI/FRP.
- Evaluator Agent: Reads serial/SSH outputs, analyzes latency and OOM errors, and feeds back to the Architect for autonomous iteration.