Deep learning has triggered the rapid expansion of AI industries. Still, we have a lot of difficult problems to solve. For this purpose, we employ a co-design framework in which researchers from different background get together to optimize the system architecture and algorithms simultaneously for deep learning. We believe this co-design framework gives us the next breakthrough to AI research. Our target for this three years is a real-time video analysis system for realizing safe and smart society. Its typical applications include human behavior prediction for auto driving and for surveillance.