Constant inbox checking breaks concentration, but traditional filters struggle with personal context. A message can be important because of the sender, the project, the wording, or a preference that changes over time.
Email AI Assistant is a desktop application that explores a more adaptable approach while keeping model processing on the user's machine.
How it works
The application connects to Gmail with user-approved permissions, retrieves new messages, and processes them with a local Llama model running through Ollama. The model evaluates each message against the user's saved preferences and decides whether a notification is warranted.
The assistant also provides a chat interface. Instead of maintaining complex rules, a user can explain why a message was or was not important. Those corrections become part of the context used for later decisions.
The data boundary
Email content can be highly sensitive. In this design, the language model runs locally rather than sending message bodies to a hosted inference API. Gmail access, local model processing, preference memory, and notifications remain separate parts of the system so each boundary is visible and testable.
What the project demonstrates
The project is less about automatically reading every message and more about building a controlled assistant around a real workflow. The user decides which account is connected, what the assistant should care about, and how its behavior changes over time.
The source code is available in the Email AI Assistant repository.