Prema Inbox Triage AI

Production-grade FastAPI + React/Vite automation platform that syncs Gmail inboxes, classifies messages with LLMs, and drafts actionable replies.

Built by Prema Vision LLC, an AI automation consultancy led by Denys Korolkov.

⭐ Live Demo

Web UI (Human-in-the-loop Dashboard) 👉 https://prema-inbox-triage-ai.vercel.app

API (Swagger / OpenAPI) 👉 https://prema-inbox-triage-ai.onrender.com/docs

Demo Video (2 min walkthrough) 👉 https://youtu.be/OHVMQciu70U

Short, real-world demo showing:

• Inbox sync (Mock + Real Gmail modes)
• AI classification (Sales / Support / Noise)
• Context-aware reply drafting
• Human review & approval before sending

⭐ Walkthrough

---

Elevator Pitch

Inbox Triage AI solves the "buried lead" problem by converting chaotic email streams into structured, actionable workflows. The system ingests messages, applies LLM classification (Lead, Support, Spam), and auto-drafts context-aware replies.

Operators manage the process via a dedicated "human-in-the-loop" dashboard, enabling review and approval in seconds—combining automation speed with operational control.

Why This Project Matters

For businesses drowning in inbound communications, this architecture demonstrates a scalable path to automation:

• Sales Ops: Instantly flag high-value leads and draft initial outreach, reducing speed-to-lead time.
• Support Ops: Auto-triage tickets and draft responses to common queries, freeing agents for complex issues.
• Engineering Quality: Shows how to build AI agents that are testable, observable, and reliable—moving beyond fragile scripts to production-grade software.

Core Features

• Dual-Mode Sync Engine: Seamlessly switches between real Gmail OAuth integration and a rich Mock Generator for safe demos and testing.
• AI Classification Pipeline: Uses OpenAI (GPT-4o) to categorize emails by intent, priority, and sentiment, while extracting key entities.
• Context-Aware Drafting: Automatically generates reply drafts based on email context and historical patterns.
• Operational Dashboard: Dedicated React/Vite dashboard for reviewing classifications, editing drafts, and managing inbox state.
• Robust Persistence: SQLModel (SQLite) backend ensures data consistency, audit trails, and easy state management.
• Developer Experience: Fully typed codebase (Python/TypeScript), dependency injection, and comprehensive E2E testing (Playwright).

⭐ Screenshots

Desktop View

Complete dashboard showing inbox summary with AI-powered categorization, sales lead detection, and reply drafts

Mobile View

Fully responsive mobile interface with touch-optimized controls and summary cards

Architecture Overview

The system follows a domain-driven design with a clear separation of concerns. A FastAPI backend orchestrates the ingestion, classification, and response workflows, identifying clear boundaries between external providers (Gmail, OpenAI) and core business logic.

graph TD
    Client[React Frontend]
    
    subgraph Backend [FastAPI Backend]
        API[API Routes]
        
        subgraph Services
            Ingestion[Ingestion Service]
            Classification[Classification Service]
            Reply[Reply Service]
            Send[Send Service]
        end
        
        subgraph Data [Data Layer]
            Repo[Email Repository]
            DB[(SQLite DB)]
        end
        
        subgraph Providers [Provider Layer]
            Gmail[Gmail Provider]
            LLM[OpenAI Client]
        end
        
        API --> Ingestion & Classification & Reply & Send
        
        Ingestion --> Gmail
        Ingestion --> Repo
        
        Classification --> LLM
        Classification --> Repo
        
        Reply --> LLM
        Reply --> Repo
        
        Send --> Gmail
        Send --> Repo
        
        Repo --> DB
    end
    
    Client -->|HTTP/JSON| API
    
    Gmail <-->|IMAP/SMTP| ExternalGmail[Gmail API]
    LLM <-->|Completion| ExternalAI[OpenAI API]

Loading

Data Flow / AI Flow

graph LR
    subgraph Providers
        Gmail[Gmail API / Mock]
        OpenAI[OpenAI API]
    end

    subgraph Backend_Services
        Ingestion[Ingestion Service]
        Classify[Classification Service]
        Reply[Reply Service]
        Send[Send Service]
    end

    subgraph Data
        DB[(SQLite / SQLModel)]
    end

    subgraph UI
        Dashboard[React Dashboard]
    end

    Gmail -->|Fetch Emails| Ingestion
    Ingestion -->|Store Raw| DB
    DB -->|Unclassified Emails| Classify
    Classify -->|Prompt| OpenAI
    OpenAI -->|Category & Tags| Classify
    Classify -->|Update Status| DB
    
    DB -->|Classified Emails| Reply
    Reply -->|Prompt| OpenAI
    OpenAI -->|Draft Content| Reply
    Reply -->|Save Draft| DB
    
    DB -->|Sync State| Dashboard
    Dashboard -->|Review & Edit| DB
    Dashboard -->|Approve Send| Send
    Send -->|Dispatch| Gmail

Loading

Tech Stack

Layer	Technologies
Backend	Python 3.12, FastAPI, SQLModel, Pydantic, Poetry
AI & Data	OpenAI API (GPT-4o), SQLite, Vector embeddings (optional integration-ready)
Frontend	React 18, TypeScript, Vite, CSS Modules, Axios
Testing	Pytest, Playwright, Unittest
DevOps	GitHub Actions, Docker (implied), Environment Config Management

Setup & Running

1. Backend Setup

# Install dependencies
poetry install

# Configure environment
cp .env.example .env
# Edit .env to set OPENAI_API_KEY and GMAIL_USE_MOCK=True for testing

2. Frontend Setup

cd frontend
npm install

3. Run the System

# Terminal 1: Start Backend
poetry run uvicorn app.main:app --reload

# Terminal 2: Start Frontend
cd frontend
npm run dev

Access the dashboard at http://localhost:5173.

How to Use (Step-by-Step Demo Flow)

1. Sync Inbox: Click "Sync" to fetch recent emails. If in Mock mode, this generates realistic test scenarios.
2. AI Processing: Watch as the system automatically classifies emails as "Sales", "Support", or "Noise".
3. Review Drafts: Click on a "Sales" lead to see the AI-suggested reply.
4. Human Override: Edit the draft inline to add a personal touch.
5. Send: Click "Send Reply" to dispatch the email (logs to console in Mock mode, sends via API in Real mode).
6. Retriage: Use the "Retriage" button to force a re-evaluation if the context changes.

API Examples

The API is fully documented via Swagger UI at http://localhost:8000/docs.

Sync Emails:

POST /emails/sync
{
  "limit": 5
}

Get Email Details:

GET /emails/{email_id}

Send Reply:

POST /emails/{email_id}/send
{
  "body": "Updated reply content..."
}

Who This Is For

• CTOs & Founders: Evaluating build-vs-buy for internal AI operations tools.
• Enterprise Architects: Looking for reference patterns on integrating LLMs into legacy workflows safely.
• Consultants: Demonstrating the capability to deliver end-to-end AI automation, not just chat interfaces.

Extensibility & Future Enhancements

• Multi-Provider Support: Abstracted provider pattern allows easy addition of Outlook (Graph API) or HubSpot.
• RAG Integration: Connect ReplyService to a vector database to draft answers based on company knowledge bases.
• Background Workers: Move LLM processing to Celery/Redis for high-volume non-blocking throughput.
• WebSockets: Implement real-time frontend updates as emails are processed.

Contact

For collaboration, consulting, or questions: Denys Korolkov — Prema Vision LLC

📧 denys@premavision.net 🌐 https://premavision.net