Chat Automation

AI-powered chat automation platform with an Express API, Next.js frontend, and a LangGraph-based Python agent with MCP tool integrations.

Key Achievements

• Automated complex, multi-step user requests, successfully executing end-to-end task flows, by designing an agentic AI workflow engine in LangGraph that implements a Plan → Route → Execute pattern with LLM-driven task decomposition.
• Decreased overall LLM token consumption by 90% by architecting a smart integration registry that dynamically binds only contextually relevant Model Context Protocol (MCP) tools per request instead of loading the entire toolset.
• Increased autonomous task completion rate by over 25% by engineering self-correction mechanisms and reasoning loops that allow the AI agent to dynamically evaluate output and gracefully recover from tool errors.
• Guaranteed the safe execution of AI-driven actions with real-world consequences, resulting in zero unauthorized system mutations, by implementing a state-based Human-in-the-Loop (HITL) system that supports mid-workflow approve, edit, and skip decisions.
• Ensured high availability of the real-time AI workflow engine, maintaining 99.9% system uptime, by provisioning and deploying the production backend infrastructure and PostgreSQL databases on AWS EC2 instances.

Live: Web · Agent · API

Technical Blog: Building an Autonomous AI Agent: A Deep Dive into the Architecture

Architecture

Service	Tech	Port
Web	Next.js 16, React 19, TailwindCSS	3000
API	Express 5, tRPC, Passport.js, Prisma, JWE	8000
Agent	FastAPI, LangGraph, LangChain, MCP	8001
Database	PostgreSQL 16	5432
Proxy	Nginx	8080

┌──────────────────────────────────────────────────┐
│                   Nginx (:8080)                  │
│                                                  │
│   /          → Web (:3000)                       │
│   /auth/*    → API (:8000)                       │
│   /trpc/*    → API (:8000)                       │
│   /oauth/*   → API (:8000)                       │
│   /agent/*   → Agent (:8001)                     │
└──────────────────────────────────────────────────┘

---

Quick Start

Choose your setup method:

• Docker Setup — Recommended for quick start, all services containerized
• Local Development — For active development with hot reload

---

Docker Setup

Prerequisites

• Docker & Docker Compose

1. Setup Environment

cp .env.docker .env

Generate a session secret and fill in the required values:

openssl rand -base64 32   # paste this as SESSION_SECRET

Required values in .env:

Variable	Where to get it
SESSION_SECRET	Run openssl rand -base64 32
GOOGLE_CLIENT_ID	Google Cloud Console
GOOGLE_CLIENT_SECRET	Google Cloud Console
GOOGLE_API_KEY	Google AI Studio
TAVILY_API_KEY	Tavily

2. Google Cloud Console Setup

1. Go to APIs & Services → Credentials
2. Create or select an OAuth 2.0 Client ID (Application type: Web application)
3. Under Authorized redirect URIs, add:

   http://localhost:8080/auth/google/callback
   

4. Under Authorized JavaScript origins, add:

   http://localhost:8080
   

5. Enable these APIs in API Library:
   • Google People API (for login profile)
   • Gmail API (if using Gmail integration)
   • Google Drive API (if using Drive/Docs/Sheets/Slides)
   • Google Calendar API (if using Calendar integration)

3. Notion Setup (Optional)

1. Go to Notion Integrations → New integration
2. Set Redirect URI to:

   http://localhost:8080/oauth/notion/callback
   

3. Add to .env:

   NOTION_CLIENT_ID=your-notion-client-id
   NOTION_CLIENT_SECRET=your-notion-client-secret

4. Start

docker compose up

Access at http://localhost:8080

---

Local Development

Prerequisites

• Node.js 20+
• pnpm 9+
• Python 3.11+ with uv
• PostgreSQL 16 (local or cloud)

1. Setup Environment

cp .env.example .env

Fill in the required values in .env:

Variable	Where to get it
DATABASE_URL	PostgreSQL connection string (local or cloud like Neon, Supabase)
SESSION_SECRET	Run openssl rand -base64 32
GOOGLE_CLIENT_ID	Google Cloud Console
GOOGLE_CLIENT_SECRET	Google Cloud Console
GOOGLE_API_KEY	Google AI Studio
TAVILY_API_KEY	Tavily

2. Google Cloud Console Setup

Same as Docker Setup, but use these URLs:

• Authorized redirect URIs: http://localhost:8000/auth/google/callback
• Authorized JavaScript origins: http://localhost:3000

3. Install Dependencies

# Install Node.js dependencies
pnpm install

# Install Python dependencies
cd apps/agent && uv sync && cd ../..

4. Setup Database

pnpm --filter @workspace/database db:push

5. Start Development Servers

pnpm dev

This starts:

• Web at http://localhost:3000
• API at http://localhost:8000
• Agent at http://localhost:8001

---

API Endpoints

Authentication (/auth/*)

Endpoint	Method	Description
/auth/google	GET	Redirect to Google OAuth
/auth/google/callback	GET	Handle Google OAuth callback
/auth/logout	POST	Destroy session & clear cookies
/auth/me	GET	Get current user (requires auth)
/auth/status	GET	Check authentication status

OAuth Integrations (/oauth/*)

Endpoint	Method	Description
/oauth/gmail	GET	Redirect to Gmail OAuth
/oauth/gmail/callback	GET	Handle Gmail OAuth callback
/oauth/google-docs	GET	Redirect to Google Docs OAuth
/oauth/google-docs/callback	GET	Handle Google Docs OAuth callback
/oauth/google-sheets	GET	Redirect to Google Sheets OAuth
/oauth/google-sheets/callback	GET	Handle Google Sheets OAuth callback
/oauth/google-slides	GET	Redirect to Google Slides OAuth
/oauth/google-slides/callback	GET	Handle Google Slides OAuth callback
/oauth/google-drive	GET	Redirect to Google Drive OAuth
/oauth/google-drive/callback	GET	Handle Google Drive OAuth callback
/oauth/google-calendar	GET	Redirect to Google Calendar OAuth
/oauth/google-calendar/callback	GET	Handle Google Calendar OAuth callback
/oauth/notion	GET	Redirect to Notion OAuth
/oauth/notion/callback	GET	Handle Notion OAuth callback
/oauth/vercel	GET	Redirect to Vercel OAuth
/oauth/vercel/callback	GET	Handle Vercel OAuth callback

tRPC (/trpc/*)

All tRPC procedures are available at /trpc/[procedure].

See packages/trpc/src/routers/ for available procedures.

Agent (/agent/*)

Endpoint	Method	Description
/agent/health	GET	Health check
/agent/chat	POST	Send chat message
/agent/chat/stream	POST	Send chat message (SSE stream)
/agent/chat/status/{thread_id}	GET	Get chat thread status
/agent/chat/retry	POST	Retry failed message
/agent/chat/resume	POST	Resume interrupted chat
/agent/sync-gmail-credentials	POST	Sync Gmail OAuth credentials

---

Project Structure

chat-automation/
├── apps/
│   ├── api/                  # Express API server
│   ├── web/                  # Next.js frontend
│   └── agent/                # Python AI agent (FastAPI + LangGraph)
├── packages/
│   ├── database/             # Prisma schema & client
│   ├── trpc/                 # Shared tRPC routers & adapters
│   ├── ui/                   # Shared UI components
│   ├── eslint-config/
│   └── typescript-config/
├── nginx/                    # Nginx reverse proxy config
├── docker-compose.yml
├── turbo.json
└── pnpm-workspace.yaml

---

AI Agent Architecture

LangGraph workflow implementing Plan → Route → Execute (Auto/Approval) → Loop pattern with Human-in-the-Loop (HITL) support.

        +----------+
        |  START   |
        +----+-----+
             |
             v
    +-----------------+
    |  SMART ROUTER   |  <-- Classify integrations, check auth
    +--------+--------+
             |
             v
    +-----------------+
    |     PLANNER     |  <-- LLM creates plan with HITL flags
    | (structured out)|
    +--------+--------+
             |
             v
    +--------------------+
    |   ROUTE_EXECUTOR   |  <-- Routes based on requires_human_approval
    +--------+-----------+
             |
             v
    +--------+-------------------------+
    |                                  |
    | approval=false                   | approval=true
    |                                  |
    v                                  v
+----------+                +------------------------+
| EXECUTOR |<---+           | EXECUTOR_WITH_APPROVAL |<---+
| (auto)   |    |           | (state-based HITL)     |    |
+----+-----+    |           +-----------+------------+    |
     |          |                       |                 |
     | should_  |                       | should_         |
     | continue |                       | continue        |
     v          |                       v                 |
+--------+      |                  +--------+             |
| TOOLS  |------+                  | TOOLS  |-------------+
+--------+ route_after_tools       +--------+ route_after_tools
     | (no more tool calls)            |
     v                                 v
    +-------------------+
    |   STEP_COMPLETE   |  <-- Extract artifacts, advance index
    +---------+---------+
              |
              v
    +---------------------+
    | should_execute_next |
    +---------+-----------+
              |
      +-------+-------+
      |               |
  route_executor     end
      |               |
      v               v
    (loop)         +-----+
                   | END |
                   +-----+

Workflow Nodes

Node	Description
smart_router	Dynamic integration loading, auth pre-flight
planner	LLM creates structured plan with HITL flags
executor	Auto-execution (no approval needed)
executor_with_approval	State-based Human-in-the-Loop execution
tools	MCP tool calling (multi-hop supported)
step_complete	Clears executor state, prepares next step

Key Features

• Multi-hop tool calling: Executor can call tools multiple times
• HITL (Human-in-the-Loop): LLM decides which steps need approval
• State-based approval: Approval state persists across tool calls
• Smart routing: Dynamic integration loading based on user's connected services
• Checkpointing: MemorySaver for workflow state persistence

---

License

MIT