Allganize AI Messenger SDK
I built an internal system that keeps AI model training history as versions, and an AI messenger SDK that takes each client's requests without touching the core code.
3SDK layers: core logic, feature API, and interface
- 2022.09 ~ 2023.03
- Allganize Korea
- SDK layer design and training management features
Core value
Each client's design and feature requests can now be met while the core code stays untouched.
Outcomes
- Every training run keeps its changes as a version, and a bad run rolls back to the previous one
- Core logic, feature API, and interface split into three layers, each extended on its own
- Live notifications show how far training has progressed
Tech stack
Context
I built an internal system that manages AI model modification, training, and generation by version, and an AI Messenger SDK distributed to external companies. The internal system had to let people track and control training in real time. The SDK had to absorb design and feature requirements that differed from client to client, layer by layer.
Decisions and implementation
- AI model training system: I built version management for each model modification, training, and generation process.
- I used GraphQL subscriptions to send real-time notifications of training progress.
- AI Messenger SDK: I designed a modular three-layer architecture (Core, Headless API, and UI) so each feature could be extended independently.
- I built a CSS isolation environment on top of iframe isolation.