Dependency Injection for On-Device AI Systems
summary
Architecture injects hardware, runtime and model details via JSON so apps run unchanged while backends swap under the hood. Implementation uses NNStreamer plugins (new backend in <800 lines) and Tizen packaging to ship model packages separately from apps, supporting on-device version tags, rollbacks and sharing. Results: support for 37 food types in refrigerators, seven-year platform updates, and model release cycles cut from 8–12 weeks to 2–3 weeks; all code upstreamed and released in 2025.
AI-written summary · about
chapters (7)
- 0:00On-device AI systems intro
- 1:49On-device AI challenges
- 7:08Dependency architecture approaches
- 12:45Configurable ML APIs
- 19:51Hardware abstraction & support
- 23:21Open-source AI products
- 28:14Summary & call to action
related talks
35mNNStreamer: Efficient and Agile On-Device AI Application Development with Neural-Net Pipelines
01mWhy edge AI is becoming the default for devices
24mWhat does AI mean for IoT development?
01h 02mArchitecture Patterns for AI-Powered Applications
32mBuilding Once, Running Everywhere: The Future of Portable AI Applications
44mArtificial Intelligence in Small Embedded Systems
19mDeveloper Centric AI
30mAccelerating AI Application: NAVER's Journey from Model to Application