jev-chat: Uses Jev model to assess counterpart intent and inspect LLM reply candidates

Details

An Android chat assistant app that uses TypeSafe's Jev decision model to determine counterpart intent and risk level, inspects reply candidates generated by an LLM (such as DeepSeek) with Jev, and fills only the input field
There is no code path for pressing the send button, so the user sends the message after final confirmation; Jev is a System One model that answers with probabilities for yes/no, multiple-choice, and step-by-step scores instead of generating sentences
The global version (v0.1.0) supports WhatsApp 1:1 conversations and recorded 48/48 for action detection, 51/54 for negative behavior, 73/74 for conflict, and 37/37 for tone in an evaluation based on 37 examples
The Chinese version (v1.7) supports QQ, X, Feishu, and WhatsApp, with Jev taking approximately 900ms, 1,000 tokens, and $0.00004 for 7 questions
The Windows sister project (v0.1.12) supports KakaoTalk and shortened frame processing time from 2.0 seconds to 0.20 seconds by switching to Windows.Media.Ocr
The app is free, but model invocation costs are paid using the user's API key, and the counterpart's messages are sent to OpenRouter and model providers
Released under the MIT license for free personal and commercial use, but unauthorized use of trademarks and domains such as 'Jev Chat Assistant' is prohibited

Key point