Overview

The project helped people manage their inbox and automate the grunt work of dealing with the many emails one receives every day. The goal was to build and deploy a web application using LLMs to automate tasks like grouping, labeling, drafting common responses, forwarding, and archiving. Rules could be given as natural language instructions:

• Label all receipts as “receipts” and organize into a folder
• Archive any one-time passwords or verification links after they are read
• Label emails which need a response as “Response Needed”
• Respond to cold emails with “Speak to someone else about this” and forward accordingly

General workflow. Every user-created rule triggers the creation of a corresponding LLM prompt and email action, entered into a database. Upon receiving an email, these prompts are retrieved and combined with the email’s content into an API call to a fine-tuned LLM, which responds with the relevant actions. Email retrieval and actions use Google’s Gmail API and OAuth 2.0. A good amount of work went into making this workflow manageable for every email sent to every user.

Difficulty — Beginner

Project members were not expected to have ML experience, but needed to be familiar with coding. The project required no more ML knowledge than proficiency working with LLMs — but it did require substantial time commitment relative to a member’s existing software and web development knowledge, since the goal was to build and deploy a usable product.

Milestones

October 1. Define the project stack and software requirements, set up the development environment, begin initial frontend development and data collection. Define scope, features, requirements, and stack based on member experience. Set up a dev environment and start building the frontend with basic UI components and navigation. Work on data collection for LLM fine-tuning.

November 1. Develop core backend logic and functionality (auth, database querying), fine-tune the LLM and integrate it with email handling for basic rules, continue developing frontend features. Develop backend modules for email fetching, preprocessing, classification, and actions via the Gmail API. Enhance the frontend to support rule creation and management.

December 1. Integrate the web app end to end, build support for custom rules with natural language, and test. Create a framework using mock emails and rules for testing. Debug and refine system prompts and email preprocessing based on tests. Enhance UI/UX based on feedback.

January 1. Deployment, hosting, and making the system work in production. Deploy frontend and backend after purchasing a domain. Set up automated tests and CI/CD pipelines. Collect data about incorrect classifications and use it to refine prompts and fine-tuning. Get users.

Team

Three members, with fluid responsibilities falling into three buckets: frontend development, backend development, and LLM fine-tuning / data collection / prompt testing. One in-person meeting every week, roughly an hour.