Build with AI - APL GDG Pune

Ritesh Kshatriya — Sun, 17 May 2026 13:09:19 +0000

🏏 Shipping "Captain Cool": How I Built a Multi-Agent IPL Strategist in 3 Hours
📋 The Mission
I recently attended the GDG Cloud Pune: Agentic Premier League 2026. The challenge was wild: "Vibe-code" a live solution during an IPL match that acts as an autonomous virtual captain.

The goal? Build "Captain Cool"—an agentic AI system that doesn't just predict scores, but makes tactical decisions (bowling changes, field setups, Impact Player subs) using a team of specialized AI agents.

🛠️ The "Google House" Tech Stack
To stay eligible for the APL, I went 100% Google Cloud. Here’s the engine under the hood:

Google Antigravity: My primary IDE. Using the .antigravity/ folder and agent traces allowed me to iterate faster than standard coding.

Gemini 2.5 Pro & Flash: The "brain" of each agent, utilizing Context Caching to keep the match history (every ball and wicket) persistent and cheap.

Agent Development Kit (ADK): Used to orchestrate the multi-agent debate and "SkillToolsets."

Function Calling: Real-time tool use to fetch live weather and pitch data.

🏗️ Architecture: The Brains Behind the Boundary
"Captain Cool" isn't just one prompt. It's a boardroom of four distinct Gemini-powered agents:

The Stats Analyst (Gemini 2.5 Flash): Uses a tool call to a Cricket API to pull historical "Match-ups" (e.g., How does the current batsman fare against left-arm spin?).

The Strategist (Gemini 2.5 Pro): Proposes the primary move based on game state and pitch conditions.

The Devil’s Advocate (Gemini 2.5 Pro): Challenges the Strategist. If the Strategist says "Bowl the pacer," the Advocate argues "The dew factor makes the ball slippery; stick to the off-spinner."

The Match Commentator: Translates the technical jargon into "cricket talk" for the fans.

The Multi-Turn Loop: I implemented a 3-turn reasoning chain. The Strategist proposes → The Advocate challenges → The Strategist confirms or pivots. This back-and-forth is what makes the final call feel human.

🕹️ Sample Scenario: The Death Overs
Match State: 18th Over | 28 runs needed | 4 wickets down | Dew falling heavily.

Internal Debate Traces:

Strategist: "Bring in the strike bowler for the 19th over to seal it."

Devil’s Advocate: "Wait. The pitch is two-paced and the batsman is a known power-hitter against pace. If we burn our strike bowler now and he goes for 15, we have no cushion for the 20th."

Final Call: "Give the 19th to the mystery spinner. Cramp them for room. Save the pacer for the final 6 balls."

🚀 Student Perspective & Learnings
As a first-year student, building in Google Antigravity felt like having a senior dev sitting next to me. Instead of wrestling with boilerplate, I spent my 3 hours focusing on the Agentic Design—making sure the agents actually disagreed with each other to find the best strategy.

Key Takeaway:
Agentic AI isn't about the LLM being "smart"; it's about the system architecture being robust enough to handle conflicting information.

🔗 Resources
GitHub Repo: Link to your Public Repo

AI Studio Prompt: [Link to your shared Prompt]

Live Demo: [Link to hosted App]

GDGCloudPune #AgenticPremierLeague #BuildWithAI #GoogleCloud #GeminiAI #VibeCoding #StudentDeveloper

Final Check for Submission:
[ ] Repository is Public?

[ ] .antigravity/ folder is committed?

[ ] No OpenAI/Claude references in code?

[ ] Logic for the "Devil's Advocate" is visible in the UI?

Building the future of sports tech, one over at a time! 🏏🔥

DEV Community: Ritesh Kshatriya

Build with AI - APL GDG Pune

GDGCloudPune #AgenticPremierLeague #BuildWithAI #GoogleCloud #GeminiAI #VibeCoding #StudentDeveloper