Building “Captain Cool AI” — A Multi-Agent IPL Strategist Using Google Gemini

Mainak Dasgupta — Sun, 17 May 2026 12:34:17 +0000

Cricket has always been called a captain’s game.

A single bowling change, a field adjustment, or the timing of an Impact Player can completely shift the momentum of a match. During the IPL, captains like Dhoni, Rohit Sharma, and Hardik Pandya constantly make tactical micro-decisions under pressure.

So I thought:

What if an AI system could think like an IPL captain?

That idea became Captain Cool AI — a multi-agent tactical cricket strategist powered entirely by the Google Gemini ecosystem.

Built for a Google-focused hackathon, the goal was not to create just another chatbot, but to build an agentic AI system where multiple AI personalities debate and collaborate before making a final cricketing decision.

🚀 The Core Idea

The system takes a live IPL match situation as input:

Current score
Overs
Wickets
Batters at crease
Pitch condition
Dew factor
Match situation

…and generates:

✅ The next tactical move
✅ The cricket reasoning behind it
✅ Internal debate between multiple AI agents
✅ Final captain-style decision

Instead of a single LLM response, the system behaves like a real IPL dugout discussion.

🧠 Why Multi-Agent Instead of One Prompt?

One of the hackathon requirements was building a truly “agentic” system.

That means:

multiple independent agents
different responsibilities
actual collaboration
visible disagreement

So instead of:

“One AI pretending to be many”

I created separate Gemini-powered agents with unique personalities and responsibilities.

🏗️ System Architecture

User Input
   ↓
Match Context Parser
   ↓
Strategist Agent
   ↓
Devil’s Advocate Agent
   ↓
Stats Analyst Agent
   ↓
Strategist Revision
   ↓
Final Tactical Decision

🤖 The Agents

1. Strategist Agent 🧠

This agent behaves like an IPL captain.

Responsibilities:

bowling changes
field setup
pressure tactics
over management
batter matchups

Example reasoning:

“Bring the leg-spinner now before the left-hander settles.”

2. Devil’s Advocate Agent ⚠️

This is my favorite part of the system.

The Devil’s Advocate intentionally challenges the strategist.

Responsibilities:

expose tactical risks
challenge assumptions
suggest alternative plans

Example:

“Bad idea. Dew is increasing and gripping the ball will become difficult.”

This creates a real tactical debate instead of a generic AI answer.

3. Stats Analyst Agent 📊

This agent provides data-driven insights.

Responsibilities:

matchup analysis
venue behavior
win probability
tactical support

Example:

“Bumrah’s economy in death overs at Wankhede this season is 7.1”

⚙️ Tech Stack

The project is built entirely on Google’s AI ecosystem.

AI Stack

Google Gemini 2.5 Pro
Google AI Studio
Google Antigravity

Frontend

Next.js
Tailwind CSS
shadcn/ui

Backend

FastAPI
Python
google-genai SDK

🎨 UI Philosophy

I wanted the interface to feel like:

an IPL dugout
a tactical war room
a broadcast analytics dashboard

The UI includes:

match state input
live agent debate panels
final tactical recommendation
cricket-style reasoning

Instead of technical AI jargon, the responses sound like real cricket discussions.

🧩 Example Match Scenario

Input

{
  "score": "154/4",
  "over": "16.2",
  "batting_team": "CSK",
  "bowling_team": "MI",
  "striker": "Dhoni",
  "pitch": "slow",
  "dew": true
}

AI Debate

Strategist

“Bring Bumrah now before Dhoni settles.”

Devil’s Advocate

“But then who bowls the 19th over?”

Stats Analyst

“Bumrah concedes only 6.8 economy in overs 16–20.”

Final Decision

“Use one Bumrah over now and reserve another for the 19th. Attack Dhoni early before the required rate drops.”

💡 Biggest Engineering Challenge

The hardest part wasn’t calling Gemini APIs.

It was designing:

believable agent personalities
tactical disagreement
natural cricket reasoning

If all agents sound identical, the system feels fake.

So each prompt had:

a unique role
different priorities
different communication style

This made the debate feel much more realistic.

🔥 What Makes This Different?

Most AI sports projects generate:

“Based on data, expected value suggests…”

I wanted this to sound like:

dugout strategy
TV commentary
IPL tactical chatter

The goal was:

AI that feels like cricket.

📈 What’s Coming Next?

Currently, Phase 1 is complete.

Upcoming upgrades include:

real cricket API integration
Gemini function calling
win probability engine
live Cricbuzz URL parsing
memory across overs
voice commentary mode
counterfactual analysis

Example:

“If Hardik bowls instead of Bumrah, win probability drops by 8%.”

🏆 Final Thoughts

This project made me realize something interesting:

The future of AI applications is not just about smarter models.

It’s about:

collaboration
reasoning
explainability
personality
interaction design

Cricket is already a game of tactical intelligence.

Building an AI captain felt like the perfect intersection of:

sports
strategy
storytelling
agentic AI

And honestly…

Watching AI agents argue over death-over bowling plans is surprisingly entertaining 😄

GitHub

Mainak29 / captain-cool-ai

Built With ❤️ Using Google Gemini

DEV Community: Mainak Dasgupta

Building “Captain Cool AI” — A Multi-Agent IPL Strategist Using Google Gemini

🚀 The Core Idea

🧠 Why Multi-Agent Instead of One Prompt?

🏗️ System Architecture

🤖 The Agents

1. Strategist Agent 🧠

2. Devil’s Advocate Agent ⚠️

3. Stats Analyst Agent 📊

⚙️ Tech Stack

AI Stack

Frontend

Backend

🎨 UI Philosophy

🧩 Example Match Scenario

Input

AI Debate

Strategist

Devil’s Advocate

Stats Analyst

Final Decision

💡 Biggest Engineering Challenge

🔥 What Makes This Different?

📈 What’s Coming Next?

🏆 Final Thoughts

GitHub

Mainak29 / captain-cool-ai

Built With ❤️ Using Google Gemini