Building “Captain Cool”: A Multi-Agent IPL Strategist Powered by Google Gemini in 3 Hours
Cricket is a captain’s game. But what if the captain was an autonomous, multi-agent AI “Brain Room” analyzing live weather, pitch soil, and historical matchups in real-time?
This weekend, I participated in the Agentic Premier League (APL) Hackathon organized by the incredible team at #gdgcloudpune. The challenge? Build an agentic AI system that acts as a virtual IPL captain—making tactical decisions under pressure exactly like Dhoni, Rohit, or Hardik. And the golden rule: It had to be built entirely on the Google Gemini ecosystem within a grueling 3-hour window.
Here is a deep dive into how I built Captain Cool 🏆.
🌟 The Problem: Moving Beyond the “Chatbot”
When building an AI for sports strategy, a generic LLM wrapper won’t cut it. If you ask a standard chatbot, “Who should bowl the next over?”, it will give you generic advice.
Real cricket requires context: Is dew setting in? Is the wind blowing toward the short boundary? Is it a red soil or black soil pitch?
Captain Cool solves this by moving away from a single prompt and utilizing an adversarial multi-agent architecture. You input the current match state, and the system replies with:
- 🎯 The Next Decision (e.g., bowling changes, field setups, Impact Player usage).
- 🧠 The Reasoning (explained in authentic commentator language).
- ⚔️ The Internal Debate (a raw, unedited back-and-forth between specialized agents).
🏗️ Architecture: Inside the “Brain Room”
To achieve this, I used Google’s Agent Development Kit (ADK) and the @google/genai SDK to create three distinct, named agents powered by gemini-2.5-pro. They don’t just generate text; they debate.
The Agent Lineup
- 🔬 The Head Analyst: The data purist. Grounds proposals strictly in match history, player matchups, boundary dimensions, and venue run-rates. Proposes the initial tactical move based on max-probability outcomes.
- 😈 The Devil’s Advocate: The risk-taker. Contrarian mindset focused on exploiting dew factors, pitch behavior, and psychological pressure. Aggressively challenges the Analyst’s every move.
- 👑 The Virtual Captain: The pragmatic leader. Synthesizes the debate, weighs resources (like remaining overs), evaluates the game phase, and delivers the final, definitive decision.
The Debate Flow
- Match State Input 2. Analyst proposes a move.
- Devil’s Advocate challenges and counters.
- Virtual Captain evaluates the conflict and makes the final call.
- The UI renders the fan-friendly commentary output alongside the raw debate log.
🌪️ Innovation: Real-Time Agentic Tooling
An AI is only as good as its data. To elevate Captain Cool, I equipped the agents with real-time API tools using Gemini Native Function Calling:
- Micro-Climate & Dew Prediction Engine: The Devil’s Advocate uses a Weather API to fetch live dew points, humidity, and wind speed. If humidity crosses 75% at a coastal venue like Chepauk, the agent dynamically calculates a “grip loss percentage” and forces the Captain to hold back spin bowlers due to a wet ball.
- Geospatial Stadium & Pitch Analytics: The Head Analyst uses a venue database to map boundary dimensions and pitch soil profiles. If the wind is blowing towards a short 60m boundary at Chinnaswamy, the agent explicitly restricts bowling into the wind and suggests specific field placements (like a deep point) to protect the vulnerability.
These tools guarantee the debate is mathematically grounded in the actual physics of the current match.
🔒 The Stack: Dual-Layer Auth & Modern Web
The platform isn’t just a backend script; it’s a fully realized web application with a sleek, monochromatic glassmorphism UI.
Authentication: I implemented a Dual-Layer Auth Gateway converging Web3 and Web2:
- 🦊 Web3: MetaMask (
window.ethereum) handshake for decentralized identity. - 🔥 Web2: Firebase Auth for classic Email/Password access.
Users configure their own Gemini API key via a secure settings dashboard, ensuring keys are safely held in the application state and never logged to the client console.
Tech Stack Breakdown:
- Frontend: React 19, Vite 8, Tailwind CSS 4, Framer Motion
- Backend: Hono.js on Bun (for blistering fast API routes)
- AI Orchestration: Google Gemini 2.5 Pro, Agent Development Kit (ADK)
- IDE: Google Antigravity (Vibe-coded the entire session here!)
🎮 A Live Scenario Walkthrough
What does this look like in action? Let’s say you input this state into the Brain Room:
Innings 2, Over 15.2, 42 runs needed off 28 balls. Big hitter on strike. Left-arm spinner has 1 over left. Dew is actively setting in. Venue: Wankhede.
The Debate Visualizer outputs:
🔬 The Head Analyst: “Brings on the left-arm orthodox spinner. Matchup data shows the batsman struggles with away spin.”
😈 The Devil’s Advocate: “Object. The ball is soaking wet due to heavy dew. The spinner will slip, lose control of length, and release pressure. Bring on the express pacer for cross-seam deliveries instead.”
👑 The Virtual Captain: “Debate closed. We save the spinner for the long-boundary side later. The pacer bowls over the wicket, targeting hard lengths into the pitch.”
🏅 Final Thoughts & Shoutouts
Building a multi-agent orchestrated system in just 3 hours was an absolute adrenaline rush. It proved to me that the future of AI development isn’t just about crafting the perfect prompt; it’s about building ecosystems where specialized AI models can reason, argue, and utilize tools autonomously.
A massive shoutout to #gdgcloudpune, especially @antrixsh_gupta **, **Pratik Kale, and the entire organizing team. The problem statement was brilliant, the vibe was electric, and events like this push the developer community to level up in incredible ways.
Cricket is a captain’s game. Now AI captains it too. 🏆
👇 Check out the code and try it yourself:
- GitHub Repository: Precise-Goals/capcool
- Tech: Google Gemini, React, Bun, Next.js, Firebase, Web3
Let me know in the comments what you think of multi-agent architectures or how you would have set the field for the final over! 🏏🔥
