---

🧠 Building a Multi-Agent AI Calendar Assistant with GenAI + Google Calendar

date: 2025-04-19

#genai #langchain #googlecalendar # agents

"What’s on my calendar today?"
"Can you schedule a sync tomorrow at 2pm?"
"Cancel the 4 PM AI meeting."

These aren't commands typed into a complex UI — they're natural phrases spoken or typed to a multi-agent AI assistant built using Gemini LLM, LangChain, and Google Calendar API. This blog is the story of how I built such an assistant for the Google GenAI Capstone, and the technical decisions that shaped it.

---

🎯 The Problem

Managing a calendar is surprisingly tedious — hopping between tabs, scanning through dates, clicking on time slots, and manually updating events. It's a cognitive tax we’ve all accepted for years.

With GenAI tools at our fingertips, I asked:
“Why can’t I just talk to my calendar like I would to a human assistant?”

---

💡 The Vision

I imagined an assistant that could understand natural queries like:a

• "List my events today"

• "Schedule a project sync on Sunday at 11am"

• "Delete the AI demo"

• "Reschedule the design review to next Friday"

But to truly pull this off, it needed more than just an LLM. It needed orchestration, understanding, memory, and integration.

---

🧠 Architecture Overview

At the heart of the assistant lies a planner → executor pipeline, built using LangChain’s agents and Google Calendar APIs.

---

🧱 Core Components

Each piece of the pipeline was built with purpose. Here’s why they exist:

1. 🕵️‍♂️ Intent Classifier (intent_chain)

Before doing anything, the assistant checks if your message is even calendar-related.

Why?
Users often say “hello” or “what’s up”. Rather than returning a confusing error or trying to parse it as a command, we route it to a small talk model.

intent = intent_chain.invoke({"user_input": user_input})
if "no" in intent.content.lower():
    reply = small_talk_chain.invoke({"user_input": user_input})
    return reply.content.strip()

...

if action_type == "create_event":
    return scheduler_agent(parsed)
elif action_type == "list_events":
    return get_upcoming_events()
elif action_type == "delete_event":
    return delete_event(parsed.get("summary", ""))

This keeps the assistant friendly, and allows casual conversation.

---

2. 🗺️ Planner Agent (planner_chain)

This is the brains of the operation.

It takes user input like:

“Schedule a team sync tomorrow at 2 for one hour”

And returns structured JSON like:

{
  "action_type": "create_event",
  "summary": "Team Sync",
  "start_time": "2025-04-20T14:00:00",
  "end_time": "2025-04-20T15:00:00",
  "timezone": "Asia/Kolkata"
}

This approach uses structured output generation with Gemini Flash, allowing the assistant to plan what needs to be done — not just answer textually.

---

3. ⚙️ Executor Tools

Depending on the "action_type", the planner’s output is routed to the right handler:

• list_events → returns a list of upcoming events

• create_event → calls a scheduler agent to create a new event

• delete_event → deletes matching events

• update_event → placeholder for future support

This separation of planning and acting is what makes the agent robust.

---

💬 Real Conversation Flow

Here’s an actual interaction from the assistant:

📝 Ask me anything: list events
🤖 You have 3 events today:
- 12:00 PM - Design Sync
- 4:00 PM - AI Demo
- 6:00 PM - Retro Review

📝 Ask me anything: cancel the AI Demo
🤖 Deleted event: AI Demo

---

🔬 Few-Shot Prompting & Custom Instructions

To improve reliability, I fine-tuned the planner prompt using few-shot examples of how to respond to vague instructions, non-calendar prompts, and ambiguous language like “move that meeting.”

---

⚠️ Limitations

• The LLM sometimes hallucinates event names

• Doesn’t yet handle recurring events

• No fuzzy matching on typos (e.g., “AI deemo” won't match “AI demo”)

• Deletion requires exact title match

• Needs manual auth step for credentials

---

🔮 Future Possibilities

• Add memory for context-aware replies

• Use fuzzy search for smarter deletion

• Voice input via browser or mobile

• Calendar visualization

• Slack or WhatsApp integration

---

🧾 Final Thoughts

This assistant was built entirely in a Kaggle notebook, using:

• 🧠 Gemini Flash (via langchain-google-genai)

• 🔗 LangChain agents

• 📆 Google Calendar API

• 🤖 Custom planner/scheduler architecture

I hope this project inspires more natural, assistive workflows — where AI not only generates text, but takes action.

Thanks to Google and Kaggle for this 5day genai-intensive course and this opportunity!

---

📂 Code Notebook

📎 View the full Kaggle notebook
📸 Includes: code, prompts, architecture, and working examples.