CollapseOS

Inspiration

In high-stakes tournaments like the 2026 World Cup, matches often swing in short chaotic windows where teams suddenly concede. Coaches and analysts rarely have a single, unified view that shows when a collapse risk is rising, which players or patterns are driving it, and what tactical adjustments could reduce that risk in real time. We saw a gap between rich event data and practical, decision-ready insights. CollapseOS was built to bridge that gap-turning granular match data into actionable intelligence that helps coaches balance tactics, player load, and momentum before a collapse happens.

What it does

It visualizes collapse risk over time.
Predicts near-term concession risk using a lightweight ML model.
Flags risks spikes with change-points detection (CUMSUM logic)
Provides tactical summaries such as possession, pass completion, and key passes.
Generates AI-assisted coaching recommendation in Coach mode.
Integrates 2026 World Cup context.

How we built it

For backend we used:

FastAPI for API layer
DuckDB + Pandas for fast analytical queries and precomputed features
scikit-learn for near-term collapse prediction
CUSUM-style detection to identify risk jumps
NetworkX to construct pass networks from StatsBomb event data
A reusable coach_tactics normalization layer to standardize StatsBomb JSON into consistent metrics

Frontend:

React, CSS for UI
Interactive pass networks Recharts / Plotly for time-series risk visualization
Gemini integration for AI-driven coach recommendations

Challenges we ran into

Schema alignment: StatsBomb’s event structure is powerful but complex. Mapping passes, formations, and outcomes into both DB and UI consistently required several iterations.
Fallback logic: Supporting both real and synthetic data without branching the entire codebase was non-trivial.
UX cohesion: Combining ML risk scores, tactical summaries, counterfactual simulations, and AI suggestions into one coherent workflow required careful layout and prioritization to avoid overwhelming users.

Accomplishments that we're proud of

A single, unified coaching dashboard that connects predictive risk, tactical context, and player impact.
Real integration with StatsBomb event data while remaining fully demo-ready with synthetic data.
A reusable tactical normalization layer that powers both analytics and UI consistently.

What we learned

Normalizing event data once and reusing it across the API and frontend dramatically improves maintainability.
DuckDB + Pandas is surprisingly powerful for hackathon-scale analytics.
A single stable “coach view” API endpoint simplifies frontend logic and increases reliability.
Predictive models are most useful when embedded directly into a decision workflow—not shown as isolated metrics.

What's next for CollapseOS

Improve the collapse-risk model with richer features and labeled training data.
Expand to additional competitions and leagues.
Deepen tactical metrics (pressure events, progressive passes, shot creation chains).
Deploy publicly to make it usable without local setup.