An experiment in personal AI: understanding which emails deserve attention, with a language model running on your own device.
Importance is personal
An important email is not always from a particular address or written with a particular keyword. Its importance may depend on a project, a conversation, or something you are waiting for. Those details change over time.
Email AI Assistant explores a more personal approach. It connects to Gmail with the user’s permission, retrieves messages, and uses a local Llama model through Ollama to consider them against saved preferences.
Corrections are part of the interface
The desktop application includes a conversation with the assistant. A person can explain why a message mattered, or why an interruption was unnecessary. That feedback becomes context for later decisions.
This changes the design question. Alongside deciding whether an email is important, the system has to make it easy for a person to correct it. Preferences need to remain understandable and under the user’s control.
Keep the boundaries visible
The language model processes messages locally. Gmail access, model inference, preference memory, and notifications are separate parts of the application. That separation makes it easier to understand where data goes and which permissions are needed.
This is an experiment, rather than a claim that every inbox can be handled automatically. Its useful lesson is that a small, personal task can reveal important product questions about attention, privacy, and control.
The source is available for people who want to inspect the design or run the project themselves.