AI email assistant
A Chrome extension that helps people with busy inboxes triage, read and reply faster, without ever sending anything they haven't approved.
Chrome Extension APIJavaScriptNode.jsPythonOpenAI APISQLite
- Built an LLM pipeline for classification, priority detection, thread summaries and reply drafts.
- Designed a human-in-the-loop approval step so AI-written replies can't go out without confirmation.
- Interviewed people at large organisations and turned their feedback into requirements.
Step 1 of 5
Gmail
The extension's content script notices when an email is opened and adds a panel next to it.
Step 2 of 5
Extract
The thread is read from the page. Raw email content is only held in memory for the request and never written to disk or a database.
Step 3 of 5
Backend
A small backend sends the thread to an LLM for a priority label, a summary (under 30 words for short threads, under 100 for long ones) and a reply draft. Drafts time out after 10 seconds so the panel never hangs.
Step 4 of 5
Panel
Priority, summary and an editable draft appear in the panel. You can change anything before it goes anywhere.
Step 5 of 5
You approve
Sending is always your action, after a confirmation step that reminds you the draft was written by AI. OAuth scopes are read-only, so the extension can't send, delete or change mail on its own.
Try it
From Lab Operations
Your building access card expires this Friday
Hi Lawrence, our records show your access card for the engineering building expires on Friday. To keep access, please renew it at the service desk before then, bringing your student ID…
- PriorityNeeds action this week
- SummaryAccess card expires Friday. Renew at the service desk with your student ID.
- Draft reply
This reply was written by AI. Send it as it is?
A mock of the approval step, so nothing here is actually sent. The real extension does the classifying, summarising and drafting with an LLM.
The problem
People in large organisations lose a lot of their day to email. We interviewed some of them and turned their feedback into our requirements: priority detection (what's urgent?), concise summaries (what's the short version?) and hands-free interaction.
The rules we built around
| Area | Rule |
|---|---|
| Permissions | Read-only OAuth scopes. The extension never asks to send, delete or modify mail. |
| Privacy | Raw email content is transient. It's never stored on disk or in a database. |
| Human in the loop | Sending is always an explicit action after a confirmation step. |
| Speed | Draft generation times out after 10 seconds. |
| Summaries | Under 30 words for short threads, under 100 for long ones. |
| Accuracy | Priority classification should reach over 85% F1. |
The approval step matters most. Language models sometimes write things that are confidently wrong or just off in tone, and an email that's already sent can't be taken back.
Where it's at
The panel, Gmail extraction, the confirmation step and the LLM-backed drafting, summarising and classifying work. Next: real Google and Microsoft sign-in, voice commands with speech recognition, and Outlook support.
History
-
Gmail extraction, the in-page panel, the confirmation step and LLM drafting, summarising and classifying all working end to end.
-
Started as a five-person university project. Interviewed people who manage high-volume inboxes at large organisations.