ScoutAI

An AI tactical co-pilot for World Cup 2026 coaches, analyzes real player data in MongoDB to flag weak links, score matchups, and recommend the best formation and substitutions.

Comment

Inspiration

Every football coach faces the same problem before a big match: hours of manual analysis, spreadsheets full of stats, and gut-feel decisions that could cost a tournament. We asked ourselves — what if a coach could simply describe their concern in natural language and get an instant, data-driven answer? Scout AI was born from that question, right as the FIFA World Cup 2026 kicks off with a record 48 teams and 1,258 players.

What it does

Scout AI is an intelligent coach assistant that analyzes your starting XI against a specific opponent and tells you exactly where you're vulnerable. You select your team, your opponent, and your formation. The agent then evaluates every player's form and role fit, runs matchup analysis across key zones (your RW vs their LB, your CB vs their ST), ranks alternative formations by score, and recommends the best replacement for your weakest link — all backed by real World Cup 2026 data. A Gemini-generated coach briefing wraps everything into a concise tactical report.

How we built it

The data layer is MongoDB Atlas, powered by the BALLDONTLIE FIFA World Cup API which gave us all 1,258 players across 48 teams with real stats. We used Google's gemini-embedding-001 model to generate 3072-dimensional vector embeddings for each player profile, stored in Atlas Vector Search for semantic similarity queries. The agent itself is built on Google ADK with Gemini 2.0 Flash, connected to MongoDB through the official MCP server. The frontend is a Streamlit app that exposes the full pipeline — lineup evaluation, matchup analysis, formation optimization, and replacement recommendations — in a clean coach-facing interface.

Challenges we ran into

Getting the vector search and full-text search indexes working correctly on the M0 free tier took significant iteration. The BALLDONTLIE API's trial window of 48 hours meant we had to fetch and store all 1,258 players quickly before the window closed. Designing the matchup scoring algorithm — comparing players across positions using heterogeneous stats — required careful schema design in MongoDB to make aggregation pipelines fast and accurate.

Accomplishments that we're proud of

We built a fully functional AI agent that gives coaches genuinely useful tactical insights, grounded entirely in real World Cup 2026 data. The MongoDB Atlas integration is deep — we use Vector Search for semantic player queries, Full-Text Search for name/position lookup, aggregation pipelines for stats ranking, and persistent session memory for the agent's context. Everything runs on free tier infrastructure.

What we learned

MongoDB Atlas Vector Search is remarkably powerful for sports analytics — the ability to query "find me a player similar to Pedri but faster" using semantic embeddings opens up entirely new ways to think about player scouting. We also learned how to design agent memory that persists across sessions using MongoDB as the backing store, which makes the coach experience feel genuinely personalized over time.

What's next for Scout AI

Live match stat integration as games are played, a visual formation builder with drag-and-drop player cards, opposition scouting reports generated automatically before each match, and expansion beyond the World Cup to club football leagues.

Built With

Share this project:

Updates