jev-chat-jarvis: Identify Chat Partner Intent and Suggest 3 Reply Candidates by Reading the Screen

This app, running on Android devices, analyzes open chat windows in real time. After assessing the intent and risk level of the other party's messages, it generates three context-appropriate reply candidates and displays them on the screen. Users can manually select and fill these into the input field; there is no auto-send feature.

Unlike existing bots that require API or account access, this system uses only the system's accessibility services to read text visible on the screen. It has been verified on WeChat, QQ, X, Feishu, and others, falling back to ML Kit-based offline OCR when text extraction is difficult. No data is sent to external servers; all processing occurs on the device.

It utilizes user-registered contact information or notes as analytical context to improve reply accuracy. While it can integrate with various LLMs such as DeepSeek, the judgment logic uses a proprietary lightweight model to optimize response speed. The architecture allows support for new chat apps by simply adding adapter code.