EmailTriager

Automatically generates contextually appropriate email reply drafts by intercepting incoming messages, extracting semantic content and tone, running inference through a language model (likely Claude or GPT), and surfacing draft responses without requiring user action. The system operates asynchronously in the background, monitoring the email inbox and triggering draft generation on new messages without blocking the user's workflow.

Parses incoming email messages to extract semantic intent, urgency level, required action type (question, request, complaint, FYI), and implicit context clues (sender role, domain, previous relationship signals). Uses NLP or embedding-based classification to categorize message type and determine appropriate response strategy before draft generation, enabling more targeted reply suggestions.

Establishes persistent connection to user's email provider (Gmail, Outlook, etc.) via OAuth 2.0 or IMAP/SMTP protocols, monitors inbox for new messages in real-time or on a polling interval, and triggers draft generation pipeline automatically without user interaction. Handles authentication refresh, credential storage, and multi-account support if applicable.

Surfaces AI-generated email drafts in a user-facing interface (likely email client sidebar, dashboard, or notification) with clear visual distinction from original message. Enables user to review, edit, approve, or discard each draft with minimal friction — typically one-click send or keyboard shortcut. May include diff view showing changes from original intent or confidence indicators.

Analyzes sender metadata (domain, title if available, previous email history) and email content tone to generate replies that match the formality level and communication style of the incoming message. For example, casual Slack-style emails receive casual replies; formal corporate emails receive formal replies. Uses embeddings or fine-tuned models to capture stylistic patterns and apply them to generated drafts.

Detects the language of incoming email and generates replies in the same language, supporting at least 10-20 major languages (English, Spanish, French, German, Mandarin, Japanese, etc.). Uses language detection on input and language-specific generation models or multilingual LLM to produce grammatically correct and culturally appropriate replies without requiring user language selection.

Assigns a quality or confidence score to each generated draft (e.g., 1-5 stars, percentage confidence, or categorical labels like 'high confidence', 'review recommended') based on factors like semantic coherence, tone match, factual accuracy (if verifiable), and alignment with detected email intent. Surfaces this score in the UI to help users prioritize which drafts to review carefully vs. approve quickly.

Retrieves previous emails in the same thread or conversation chain and incorporates relevant context into draft generation. Uses vector embeddings or BM25 search to find related messages, extracts key facts/decisions from prior emails, and injects this context into the LLM prompt to generate more coherent and factually consistent replies. May include summarization of long threads to fit within token limits.