Skip to main content

Building the AI Email Assistant: Taming Inbox Overload

· 3 min read
Uday Tamma
Uday Tamma
Building AI-Powered Applications

The AI Email Assistant tackles a universal problem: email overload. Built with Google Gemini AI, it automatically categorizes, summarizes, and enables one-click actions on Gmail emails.

The Problem​

Professionals spend hours daily processing emails. Most of this work is repetitive:

  • Deciding what's urgent vs. informational
  • Deleting promotional content
  • Extracting action items from messages

The goal: reduce email processing time by 70% through AI-powered automation.

Core Features​

AI-Powered Categorization​

Gemini AI classifies emails into 6 categories:

CategoryDescriptionExample
Need-ActionRequires responseBills, important tasks
FYIInformation onlyReceipts, confirmations
NewsletterRegular digestsWeekly newsletters
MarketingPromotional contentSales, offers
SPAMUnwanted messagesSuspicious content
OtherUncategorizedEverything else

Each email also gets:

  • Subcategory (Bill-Due, Package-Tracker, JobAlert, etc.)
  • AI Summary (one sentence, ≤15 words)
  • Suggested Action (AddToCalendar, Delete, Unsubscribe, etc.)

One-Click Actions​

Category-based quick actions eliminate decision fatigue:

  • FYI/Newsletter/Marketing: Delete button (moves to trash)
  • Need-Action: Dropdown with 4 options (Add to Calendar, Add Task, Mark as Read, Delete)
  • SPAM: Unsubscribe link (auto-extracts unsubscribe email and sends opt-out)

Actions execute instantly without confirmation dialogs. The UI updates inline.

Comprehensive Logging​

Full audit trail for debugging and observability:

  • All Gemini AI interactions logged with prompts and responses
  • API calls logged with formatted Python syntax
  • Daily log rotation with search and filtering
  • Statistics dashboard (latency, operation counts)

Technical Architecture​

┌─────────────────────────────────────────────────┐
│              Web Browser                         │
│  (Email Review UI, Logs Dashboard)              │
└──────────────────────┬──────────────────────────┘
                       │ HTTP/JSON
┌──────────────────────▼──────────────────────────┐
│              Flask Web Server                    │
│  Routes: /gemini-review, /gemini-logs, /api/*   │
└──────────────────────┬──────────────────────────┘
                       │
        ┌──────────────┼──────────────┐
        │              │              │
┌───────▼─────┐ ┌──────▼──────┐ ┌────▼─────────┐
│ Gemini AI   │ │ Gmail API   │ │ Calendar/    │
│ Service     │ │ Service     │ │ Tasks APIs   │
└─────────────┘ └─────────────┘ └──────────────┘

Tech Stack​

  • Backend: Python 3.14, Flask
  • AI: Google Gemini 2.5 Flash Lite
  • APIs: Gmail API v1, Calendar API v3, Tasks API v1
  • Frontend: Vanilla HTML/CSS/JS with Jinja2 templates
  • Auth: OAuth 2.0 for Google services

Key Design Decisions​

1. No Confirmation Dialogs​

Actions execute immediately. Users processing 50+ emails don't want to click "Are you sure?" 50 times. The trade-off: occasional mistakes are easily undone (emails go to trash, not permanent delete).

2. Server-Side Rendering​

Chose Flask + Jinja2 over a SPA framework. For a personal productivity tool, simplicity wins. No build step, no framework overhead, instant page loads.

3. Comprehensive Logging​

Every Gemini interaction and API call is logged. This was critical for:

  • Debugging categorization issues
  • Understanding AI behavior patterns
  • Building confidence in the system

4. Category-Based UI​

The action controls change based on email category. Marketing emails get a simple delete button. Need-Action emails get a full dropdown. The UI adapts to the task.

Performance​

  • Single email categorization: under 2 seconds (p95)
  • Batch of 30 emails: under 60 seconds (p95)
  • Page load: under 2 seconds
  • Action execution: under 1 second

Rate limiting handles Gemini's 30 requests/minute quota gracefully.

Lessons Learned​

  1. AI summaries are surprisingly useful: The one-sentence summary often tells you everything you need without opening the email.

  2. Unsubscribe extraction is tricky: Parsing unsubscribe links from email bodies requires handling many formats.

  3. Logging saves debugging time: When categorization seems wrong, the logs show exactly what the AI saw and why it decided.

  4. Gmail API quotas matter: Batch operations and caching are essential for processing large inboxes.

What's Next​

Planned for Phase 2:

  • Smart date/time extraction for calendar events
  • Bulk actions (select multiple emails)
  • Email threading support

Future phases:

  • Natural language queries ("Show bills due this month")
  • Automation rules engine
  • Mobile app