DT42/BerryNet

Deep learning gateway on Raspberry Pi and other edge devices

What it solves

BerryNet turns a Raspberry Pi into an intelligent edge gateway capable of running deep learning models locally. It eliminates the need for cloud connectivity, saving on data transmission costs and enabling real-time responses to visual or sensor events directly on the device.

How it works

It uses a modular architecture where independent components—such as cameras, AI engines, and data collectors—communicate via an MQTT broker. The system supports two primary AI engines: LiteRT and ONNX Runtime. These engines process video frames or images using pre-trained models (like MobileNet or YOLO) and publish the results to specific MQTT topics. Users can monitor these results via a web-based dashboard or receive automated alerts through services like Telegram and Gmail.

Who it’s for

This project is for developers and hobbyists building AIoT (Artificial Intelligence of Things) applications on Raspberry Pi hardware who need a local, privacy-conscious, and distributed way to process visual data or sensor inputs without relying on external cloud services.

Highlights

  • Runs deep learning models entirely offline on edge hardware.
  • Supports diverse input sources including Raspberry Pi camera modules, USB cameras, and IP cameras.
  • Compatible with popular model formats like LiteRT and ONNX, including support for Ultralytics YOLO models.
  • Features a modular, component-based design managed by supervisor for easy configuration and scaling.
  • Includes built-in support for monitoring Zigbee sensors, such as soil moisture monitors, with automated alerting.

Related

  • Project
    pytorch/FBGEMM

    A repository of highly-optimized kernels for deep learning, providing low-precision matrix multiplication and convolution libraries for server-side inference and generative AI.

  • Project
    Xilinx/brevitas

    A PyTorch library for neural network quantization that supports both post-training quantization and quantization-aware training to optimize models for efficiency.

  • Project
    xLLM-AI/xllm

    An efficient LLM inference framework optimized for Chinese AI accelerators to enable high-throughput, low-latency enterprise deployment.

  • Project
    RightNow-AI/autokernel

    AutoKernel is an autonomous GPU kernel optimizer that takes any PyTorch model, profiles it to find slow kernels, then uses an AI agent to iteratively rewrite and benchmark Triton or CUDA C++ kernels overnight. It verifies correctness and reports the total speedup, saving developers from manual kernel tuning.