Fantastic project! Out of curiosity, do you foresee any optimization paths (like model quantization) that could lower the memory footprint down the road