jev-chat/jev-chat-jarvis

What it solves

Jev Chat Assistant acts as a "conversation co-pilot" for mobile messaging apps. It helps users understand the true intent and risk level of incoming messages and provides suggested replies, reducing the mental effort of deciding how to respond in various social or professional contexts.

How it works

The app uses Android's Accessibility Services to read chat bubbles from the screen without hooking into the app's code or accessing databases. For apps where text is not accessible in the system tree (like Lark/Feishu), it uses on-device OCR via ML Kit.

Once a message is captured, it follows a two-step AI process:

1. Analysis: A specialized model (TypeSafe Jev) analyzes the message to determine intent, risk level (1-9), and whether a response is needed.
2. Generation: A generative model (defaulting to DeepSeek) drafts three colloquial reply options, which are then ranked by the analysis model based on suitability.

These suggestions appear in a floating window, and the user can one-click fill the reply into the input box, though the app never sends the message automatically.

Who it’s for

Users of Android devices who want AI-powered assistance to navigate complex social interactions across multiple platforms like WeChat, QQ, X (Twitter), and Lark.

Highlights

• Non-intrusive: Uses accessibility services and OCR instead of API hooks or account modifications.
• Multi-platform: Supports WeChat, QQ, X, and Lark through a modular adapter system.
• Context-aware: Maintains a local knowledge base of notes and contact profiles to ensure replies are consistent with personal facts.
• Privacy-focused: API keys and chat history are stored locally on the device; content is only sent to the model during the moment of analysis.
• Human-in-the-loop: Only fills the text box; the user must manually press send.