This is Pocket, an AI assistant built by Naz Louis. The goal of this project is to run AI models completely on a Raspberry Pi 5. Since Pi 5 comes with limited resources, what he did is instead of running a large model directly on Pi 5, he made a collection of light weight models and made them work together. He designed and deployed a routing setup that routes different tasks to the right models. He picked Qwen as the LLM (Large Language Model), Fast Whisper for Speech-to-Text and Piper for Text-to-Speech. He fine-tuned function gemma on his custom data set so that he can call his custom functions.
He added two 18650 batteries to make it portable. He used 18650 battery shield for Raspberry Pi to mount the batteries. Then he added a 12 megapixel Arducam camera to the device for live visuals and a 26 TOPS (Terra Operations Per Second) Hailo AI accelerator for Edge AI processing. He used a 4.3” touch display with speaker and mic so that he can interact with Pocket. He made a custom 3D printed body and a custom UI with a lot of options. Now, he can talk to pocket, ask questions and get proper responses. He also added an option for camera which opens up the live visual from the Arducam camera and we can also run object detection with the visuals and this runs entirely on the hailo AI accelerator. The cool thing is he has made this completely open source.
