StadiumPilot

About StadiumPilot

StadiumPilot is a Generative-AI companion for FIFA World Cup 2026 match days. It serves fans and venue operators from a single application, using deterministic venue data as ground truth and Generative AI for language, reasoning, and decision support on top of it.

Challenge alignment

ThemeHow StadiumPilot addresses it
NavigationGraph-routed directions narrated by AI, with QR-style location deep links.
Crowd managementLive density heat grid, threshold alerts, and AI dispersal recommendations.
AccessibilityStep-free routing, accessible amenity finder, and a WCAG-minded interface.
TransportationMetro, shuttle, rideshare, and parking guidance combined with live crowd state.
SustainabilityRecycling and refill finder plus waste-diversion and refill KPIs.
Multilingual assistanceCopilot replies in the fan's language; PA drafts in several languages at once.
Operational intelligenceAI incident summaries, volunteer briefings, and an end-of-match report.
Real-time decision supportEach alert becomes a structured action card with human-in-the-loop confirm.

How Generative AI is used

Architecture

A typed venue graph provides routing and amenities; a seeded, deterministic crowd model turns the match minute into live density, incidents, and telemetry with no database. The AI layer (Vercel AI SDK with Google Gemini) is reached only through server routes that validate input, rate-limit per client, enforce same-origin, and cap output tokens. The fan copilot is grounded through read-only tools, so it cannot invent venue facts and prompt injection cannot cause side effects.

Accessibility

Semantic landmarks and a skip link, full keyboard operability with visible focus, ARIA live regions for streaming replies and alerts, a screen-reader table alongside the heat grid, automatic text direction for right-to-left languages, a step-free routing mode, and respect for reduced-motion preferences.

Security

Input is validated with Zod, POST routes enforce same-origin and body-size limits, each client is rate-limited, security headers are set at the edge, and the API key never reaches the browser. Model output is rendered as text with a safe markdown subset — never as raw HTML.

Built with

Next.js (App Router), React, TypeScript, Tailwind CSS, the Vercel AI SDK with Google Gemini, and Vitest. Deployed on Vercel.

Unofficial demo. Not affiliated with FIFA or any real venue. All venue data is synthetic.