mailmind

Note: update table number to 69

On average, we spend about 2 hours per day on email.

Email clients like gmail allow users to create custom filters to route their emails to different folders based on specific criterias. This feature helps users organize their inbox more efficiently and reduces the time spent searching for important messages.

However, creating these filters can be extremely time-consuming and requires a good understanding of the email client's syntax. These rules can be hard to generalize when you received hundreds of emails. Also when providers their email addresses, these rules can become outdated and need to be updated regularly.

So we tried teaching a multi-agent to help with this. We used three agents:

• Agent 1: Summarizes the activity log of the user's interactions.
• Agent 2: Responds to incoming emails and organizes them.
• Agent 3: Uploads activity to wandb and fine-tunes the model.

In order to train the agent, we built a mock email environment to simulate inbox activity and record user interactions, and have an agent learn from these interactions. And then hand it over to the agent to continue organizing when the user is afk.

After a day and a half of hacking, we were able to train the agents to organize anyone's inbox quite well.

Built With

• google
• next.js
• python
• w&b