DEV Community

Sarwadnya Mahajan
Sarwadnya Mahajan Subscriber

Posted on

# Captain Cool — Multi-Agent IPL Match Strategist 🏏

#gdgcloudpune

When the pressure is at its peak, four AI minds debate so the captain doesn't have to guess.

The Problem

IPL captaincy decisions happen in seconds. Bowl a spinner in the death overs with dew? Persist with a struggling fast bowler who has quota left? Bring in the Impact Player now or save them?

A single AI model gives you one answer. But cricket captaincy isn't about one answer — it's about considering the counter-argument, the risk, the gut instinct, and then making a call you can defend.

Captain Cool simulates this internal debate using four specialized AI agents that argue, challenge, and ultimately produce a decision that has been stress-tested before it's made.

One agent fetches live data. One makes the call like MS Dhoni. One tears it apart like Virat Kohli. One wraps it all in commentary like Harsha Bhogle. The result: a decision that's been through the fire.

Architecture Overview 

┌─────────────┐
│  USER INPUT  │  Match state, teams, venue, conditions
└──────┬──────┘
       │
       ▼
┌──────────────┐
│ ORCHESTRATOR │  5-turn debate manager + validation guards
└──────┬───────┘
       │
       ▼ Turn 1
┌──────────────────────────────────────────┐
│        STATS ANALYST (Gemini 2.5 Flash)   │
│  Tools: CricAPI live → ESPN → Fallback    │
│  + win_probability + player_stats         │
│  + venue_profile + bowler_quota           │
│  Output: Structured stats brief           │
└──────────────┬───────────────────────────┘
               │ stats_brief
               ▼ Turn 2
┌──────────────────────────────────────────┐
│        STRATEGIST (Gemini 2.5 Flash)      │
│  Personality: MS Dhoni's brain            │
│  Output: DECISION + FIELD_SETUP +         │
│          REASONING + REVISION_STATUS      │
└──────────────┬───────────────────────────┘
               │ initial_proposal
               ▼ Turn 3
┌──────────────────────────────────────────┐
│     DEVIL'S ADVOCATE (Gemini 2.5 Flash)   │
│  Personality: Virat Kohli's aggression    │
│  Output: CHALLENGE + ALTERNATIVE +        │
│          RISK_IF_WRONG + STAT_BACKING     │
└──────────────┬───────────────────────────┘
               │ devils_challenge
               ▼ Turn 4
┌──────────────────────────────────────────┐
│     STRATEGIST AGAIN (Gemini 2.5 Flash)   │
│  Receives challenge, defends or revises   │
│  Output: Final DECISION + REVISION_STATUS │
└──────────────┬───────────────────────────┘
               │ final_decision
               ▼ Turn 5
┌──────────────────────────────────────────┐
│       COMMENTATOR (Gemini 2.5 Flash)      │
│  Personality: Harsha Bhogle + Shastri     │
│  Output: 150-word broadcast commentary    │
└──────────────┬───────────────────────────┘
               │
               ▼
┌──────────────────────────────────────────┐
│          STREAMLIT UI                     │
│  5 tabs + live banner + debate log        │
└──────────────────────────────────────────┘
Enter fullscreen mode Exit fullscreen mode

Note: Replace this ASCII diagram with an Excalidraw visual before publishing.

Google Tech Stack Used

Technology Role Why Chosen
Gemini 2.5 Flash All 4 agent brains Fast, cheap, excellent at structured output and cricket reasoning
Google ADK Tool registration & agent framework FunctionTool wrapping gives judges visible tool-use in traces
Google AI Studio Prompt prototyping Tested each agent's system prompt interactively before coding
Google Antigravity Build session IDE Full project built in Antigravity — .antigravity/ folder proves it
BeautifulSoup4 / Requests Live Data Backup Used for the ESPNCricinfo web scraper when API limits are reached

Agent Profiles

1. StatsAnalyst — The Data Machine

Personality: Pure facts. No opinions. Temperature 0.1.

System Prompt:

You are StatsAnalyst — a cricket data specialist for IPL T20 matches.

YOUR ONLY JOB: Process match state data and return a structured stats brief. You are NOT a strategist. You do NOT make decisions. You report facts.

RULES:
1. You must ALWAYS report the data source (live API or hardcoded fallback) and timestamp.
2. If live data is available, use the live score and overs to override manually entered values.
3. Report batter handedness EVERY TIME — it changes field setup completely.
4. Flag any bowler who has bowled 3 or more overs with a BOWLER_QUOTA_ALERT.
5. Give win probability as a SINGLE NUMBER, not a range.
6. NEVER guess any number — only report what the tool results show.
7. Determine the match phase: powerplay (overs 1-6), middle (overs 7-15), death (overs 16-20).
Enter fullscreen mode Exit fullscreen mode

Temperature: 0.1 — Maximum factual accuracy, no creative liberty with numbers.

Input: Match state + live API data + tool results (player stats, venue, win probability)

Output Format:

LIVE_DATA_SOURCE: [source]
FETCHED_AT: [timestamp]
MATCH_PHASE: [phase]
BATTING_WIN_PROB: [X]%
PROJECTED_OR_REQUIRED: [value]
KEY_BATTER: [name] ([hand]-handed) — [stats]
KEY_BOWLER_THREAT: [name] — [reason]
VENUE_FACTOR: [insight]
BOWLER_QUOTA_ALERT: [alert or NONE]
CONFIDENCE: [level] — [reason]
Enter fullscreen mode Exit fullscreen mode

2. Strategist — Captain Cool's Brain

Personality: MS Dhoni. Ice-cold. Data-driven. Never panics. Temperature 0.4.

System Prompt:

You are the Strategist — MS Dhoni's brain. Ice-cold. Data-driven. Never panics. Makes exactly one clear call. Explains it in cricket language, not jargon.

FACTORS YOU MUST CONSIDER:
1. BATTER HANDEDNESS vs BOWLER TYPE MATCHUP
2. BOWLER QUOTA — NEVER suggest a bowler who has already bowled 4 overs
3. PHASE-SPECIFIC LOGIC — Powerplay, middle, and death require different thinking
4. DEW FACTOR — If dew is active, spinners become ineffective after over 14
5. IMPACT PLAYER — When is the right moment
6. FIELD POSITIONS — Name at least 2 specific fielding positions
Enter fullscreen mode Exit fullscreen mode

Temperature: 0.4 — Balanced reasoning, slight creativity for tactical innovation.

Input: Stats brief + match state + optional challenge from DevilsAdvocate

Output Format:

DECISION: [action]
REASONING: [3 sentences max]
CONFIDENCE: [0-100]
FIELD_SETUP: [2-3 positions]
ALTERNATIVE_CONSIDERED: [what else was considered]
REVISION_STATUS: REVISED: YES or REVISED: NO
Enter fullscreen mode Exit fullscreen mode

3. DevilsAdvocate — The Challenger

Personality: Virat Kohli's aggressive instinct. Always pushes back. Temperature 0.7.

System Prompt:

You are the DevilsAdvocate — Virat Kohli's aggressive instinct. Always pushes back. Never fully agrees. Finds the flaw in every plan. Loves to prove the obvious choice is wrong.

RULES:
1. Find at least ONE SPECIFIC weakness
2. Propose a CONCRETE ALTERNATIVE with player names
3. QUANTIFY THE RISK if Strategist is wrong
4. Back challenge with ONE SPECIFIC STAT or venue fact
5. NEVER agree completely
6. Be aggressive but LOGICAL
Enter fullscreen mode Exit fullscreen mode

Temperature: 0.7 — Creative enough to find contrarian angles, grounded enough to stay logical.

Input: Strategist's proposal + stats brief + match state

Output Format:

CHALLENGE: [specific weakness]
ALTERNATIVE: [counter-call with players]
RISK_IF_WRONG: [quantified consequence]
STAT_BACKING: [supporting evidence]
Enter fullscreen mode Exit fullscreen mode

4. Commentator — The Voice

Personality: Harsha Bhogle meets Ravi Shastri. Temperature 0.9.

System Prompt:

You are the Commentator — Harsha Bhogle's eloquence meets Ravi Shastri's passion. Speaks to fans, not to analysts. Every sentence earns its place.

RULES:
1. Use player NICKNAMES (King Kohli, Boom Boom, Captain Cool)
2. Reference VENUE CHARACTER vividly (Dharamshala = mountain fortress, Delhi = furnace)
3. Summarize dissent in one sentence starting with "The dissenting view:"
4. End with one prediction starting with "The call:"
5. MAXIMUM 150 words
6. Sound like a LIVE BROADCAST
7. NEVER use statistics — only narratives
Enter fullscreen mode Exit fullscreen mode

Temperature: 0.9 — Maximum vivid, expressive language for broadcast feel.

Input: Full debate (proposal + challenge + final decision + match state)

Output: 150-word broadcast commentary

The Tools

ADK Tool Registration

Every tool function is wrapped as a google.adk.tools.FunctionTool:

from google.adk.tools import FunctionTool

adk_get_player_stats = FunctionTool(get_player_stats)
adk_win_probability = FunctionTool(win_probability)
adk_get_live_match_data = FunctionTool(get_best_live_data)
Enter fullscreen mode Exit fullscreen mode

Only the StatsAnalyst agent receives tools. The other three agents reason purely on context.

Win Probability — DLS-Based

The tool uses a simplified Duckworth-Lewis-Stern resource table:

  • Maps wickets lost (0-9) to resources remaining at each over
  • For 1st innings: projects final score based on resources used vs total
  • For 2nd innings: compares required run rate to par RR of 9.0, adjusts for wickets in hand
  • Caps probability between 5% and 95% to avoid overconfidence 
# Simplified Win Probability Logic
par_rr = 9.0
required_rr = runs_needed / overs_left
rr_ratio = par_rr / required_rr
base_prob = rr_ratio * 50  # At par RR, 50% chance

# Adjust for wickets in hand
wicket_factor = (10 - wickets) / 10.0
batting_win_prob = base_prob * wicket_factor
Enter fullscreen mode Exit fullscreen mode

Three-Tier Live Data Fetching

This is the key differentiator for the Technical Depth rubric:

  1. CricAPI Live (Primary) — Calls /v1/cricScore and /v1/match_scorecard endpoints for real-time scores
  2. ESPNCricinfo Scrape (Backup) — Fetches and parses the match page if CricAPI quota runs out
  3. Hardcoded Fallback (Safety net) — Pre-loaded IPL 2026 season data ensures the app never crashes 
# ESPN Web Scrape Fallback Example
def get_espn_match_data(match_url):
    headers = {"User-Agent": "Mozilla/5.0..."}
    resp = requests.get(match_url, headers=headers)

    # Basic extraction from page content using regex
    score_pattern = re.findall(r'(\d+)/(\d+)\s*\((\d+\.?\d*)\s*ov\)', resp.text)
    if score_pattern:
        runs, wickets, overs = score_pattern[-1]
        return f"{runs}/{wickets}"
Enter fullscreen mode Exit fullscreen mode

The data_source field in every response tells the UI which tier was used, displayed as a color-coded banner.

The Debate Loop — Example

Scenario: PBKS vs RCB, Match 61, Dharamshala. Over 8 ball 1. RCB batting, 72/1. Virat Kohli on strike, Arshdeep Singh bowling.

Turn 1 — StatsAnalyst 

LIVE_DATA_SOURCE: hardcoded fallback
FETCHED_AT: N/A
MATCH_PHASE: middle
BATTING_WIN_PROB: 62%
PROJECTED_OR_REQUIRED: projected score 210
KEY_BATTER: Virat Kohli (right-handed) — 484 runs, avg 53.78, century last match vs KKR
KEY_BOWLER_THREAT: Arshdeep Singh — left arm swing, 13 wickets, angles into right-hander
VENUE_FACTOR: Dharamshala flat pitch, 205+ average, spin ineffective
BOWLER_QUOTA_ALERT: NONE
CONFIDENCE: high — set batter on flat pitch with short boundaries
Enter fullscreen mode Exit fullscreen mode

Turn 2 — Strategist (Initial) 

DECISION: Bring on Harpreet Brar (left arm spin) to replace Arshdeep Singh for over 9
REASONING: Kohli dominates pace in the middle overs. A left arm spinner changes the angle and forces Kohli to adjust. Arshdeep has 1 over used — save him for the death where his yorkers are gold.
CONFIDENCE: 72
FIELD_SETUP: Deep midwicket for the slog sweep, short fine leg for the sweep, extra cover to cut the drive
ALTERNATIVE_CONSIDERED: Continuing with Arshdeep — rejected because pace on this flat pitch will be dispatched by a set Kohli
REVISION_STATUS: REVISED: NO
Enter fullscreen mode Exit fullscreen mode

Turn 3 — DevilsAdvocate 

CHALLENGE: Harpreet Brar is a left arm spinner on a Dharamshala pitch where spin is NOT effective. Both matches this season went 200+ with spinners being toothless. Kohli's strike rate against left arm spin in the middle overs is elite.
ALTERNATIVE: Keep Lockie Ferguson on — his raw pace and bounce can surprise Kohli. Ferguson generates extra bounce at altitude which is harder to time.
RISK_IF_WRONG: Brar could leak 12-15 runs in the over on this flat pitch, pushing RCB's projected score past 220. That's a game-losing over.
STAT_BACKING: Dharamshala average score is 205+ and spin has been ineffective in both matches this season.
Enter fullscreen mode Exit fullscreen mode

Turn 4 — Strategist (Revised) 

DECISION: Bring on Lockie Ferguson for over 9 instead of Harpreet Brar — use pace and bounce to challenge Kohli
REASONING: The Devil's Advocate is right — spin is toothless at Dharamshala this season. Ferguson's extra bounce at altitude creates a genuine challenge even for a set Kohli. His pace makes timing harder on a quick outfield.
CONFIDENCE: 78
FIELD_SETUP: Deep square leg for the pull, fine leg back for the glance, point in a catching position
ALTERNATIVE_CONSIDERED: Harpreet Brar — rejected after Devil's Advocate correctly identified spin's ineffectiveness at this venue
REVISION_STATUS: REVISED: YES
Enter fullscreen mode Exit fullscreen mode

Turn 5 — Commentator 

High above the clouds in this mountain fortress, King Kohli is in his kingdom. Seventy-two on the board, just one down, and the maestro is painting the Dharamshala canvas with those trademark drives.

The captain had an idea — slow it down with Harpreet Brar's left arm spin. The dissenting view: "Spin at Dharamshala? Both matches here went past 200 with spinners getting carted — you need Ferguson's thunderbolts."

And the captain listened. Lockie Ferguson gets the ball. Raw pace. Extra bounce. The altitude factor.

The call: Ferguson's express pace will test even Kohli's impeccable timing, and PBKS will keep this under 200 if they strike in the next three overs.
Enter fullscreen mode Exit fullscreen mode

Live Demo Screenshots

TODO: Add screenshots of all 5 tabs and the live data banner before publishing.

  • [ ] Stats Brief tab with LIVE_DATA_SOURCE visible
  • [ ] Initial Proposal tab
  • [ ] Devil's Advocate tab
  • [ ] Final Decision tab with revision badge
  • [ ] Commentary tab
  • [ ] Live data banner (green/yellow)
  • [ ] Debate Log expander

AI Studio Prompt Links

TODO: Prototype each system prompt in aistudio.google.com, save as shareable link, paste below.

  • [ ] StatsAnalyst prompt: [link]
  • [ ] Strategist prompt: [link]
  • [ ] DevilsAdvocate prompt: [link]
  • [ ] Commentator prompt: [link]

What I Would Build With More Time

  1. Voice Input via Web Speech API + Gemini Live — Captain speaks the match state, Gemini Live processes it in real-time
  2. Real-time Cricbuzz URL Scraping — Using Gemini's URL context tool for richer live data
  3. Gemini Context Caching Across Overs — Cheap multi-turn memory so the system remembers what happened 3 overs ago
  4. Confidence Score with Counterfactual Analysis — "If we had bowled Ferguson instead of Brar, the projected score would have been..."
  5. Veo Highlight Clip Generation — Generate a short AI video clip visualizing the captain's decision moment

GitHub Repository

[https://github.com/SarwadnyaMahajan/captain-cool]

Built with Google Gemini 2.5 Flash · Google ADK · GDG Cloud Pune APL 2026

#gdgcloudpune

Top comments (0)