ReasonFlow — Autonomous Inbox AI Agent

Intelligent email processing powered by LangGraph + Gemini, with human-in-the-loop approval workflows.

Overview

ReasonFlow is a full-stack autonomous inbox agent that:

• Classifies incoming emails by intent (inquiry, meeting request, complaint, etc.)
• Retrieves relevant context from past emails, CRM data, and calendar
• Generates draft responses using Google Gemini via LangChain
• Routes low-confidence drafts to humans for approval
• Dispatches approved responses automatically

Tech Stack

Layer	Technology
Backend	FastAPI, LangGraph, LangChain + Gemini, SQLAlchemy
Frontend	Next.js 14 (App Router), shadcn/ui, Zustand, TanStack Query
Database	PostgreSQL + pgvector
LLM	Google Gemini API (via langchain-google-genai)

Architecture

┌─────────────┐     ┌──────────────────────────────────────────┐
│   Next.js   │────▶│              FastAPI Backend              │
│  Dashboard  │◀────│                                          │
└─────────────┘     │  ┌─────────┐  ┌──────────┐  ┌────────┐  │
                    │  │  Agent   │  │Retrieval │  │  APIs  │  │
                    │  │ Engine   │  │ Module   │  │        │  │
                    │  │(LangGraph│  │(pgvector)│  │Gmail   │  │
                    │  │+ Gemini) │  │          │  │Calendar│  │
                    │  └─────────┘  └──────────┘  │CRM     │  │
                    │                              └────────┘  │
                    └──────────────────────────────────────────┘
                                    │
                    ┌───────────────┼───────────────┐
                    │               │               │
               PostgreSQL        External
               + pgvector        APIs

Agent Workflow

Email Input → Classification → Context Retrieval → Decision
    → Tool Execution → Response Generation → Human Review → Dispatch

Each step is a LangGraph node with full observability and state tracking.

Quick Start

Backend

cd backend
pip install -e ".[dev]"
cp .env.example .env
# Edit .env with your credentials
alembic upgrade head
uvicorn app.main:app --reload

Frontend

cd frontend
npm install
cp .env.example .env.local
npm run dev

Docker (Full Stack)

docker compose -f docker/docker-compose.yml up --build

Project Structure

backend/
├── app/
│   ├── agent/          # LangGraph workflow engine
│   │   ├── nodes/      # Classification, retrieval, decision, generation, review
│   │   ├── state/      # Agent state definitions
│   │   └── tools/      # Tool manager + tool implementations
│   ├── api/            # FastAPI routes + middleware
│   ├── core/           # Config, database, security
│   ├── integrations/   # Gmail, Calendar, CRM clients
│   ├── llm/            # Gemini client + prompts (via LangChain)
│   ├── models/         # SQLAlchemy models
│   ├── retrieval/      # Embeddings + vector search
│   ├── schemas/        # Pydantic request/response schemas
│   └── services/       # Business logic layer
├── tests/
└── alembic/

frontend/
├── src/
│   ├── app/
│   │   ├── (auth)/     # Login, OAuth callback
│   │   └── (dashboard)/ # Inbox, Drafts, Metrics, Traces
│   ├── components/     # UI components per feature
│   ├── hooks/          # React Query hooks
│   ├── stores/         # Zustand state stores
│   ├── lib/            # API client, utilities
│   └── types/          # TypeScript interfaces

Documentation

• System Architecture
• Database Schema
• Agent Workflow
• API Reference
• Gmail Integration
• Agent Engine
• Calendar Integration
• CRM Integration
• Retrieval Module
• Inbox Dashboard
• Draft Review
• Metrics Dashboard
• Trace Viewer

Performance Targets

Metric	Target
API latency	< 300ms
Agent full cycle	< 4s
Dashboard load	< 1s

License

MIT