FIFA World Cup 2026 · Stadium Operations

NEXUSThe AI teammate that

NEXUS watches every operational feed in the control room, spots crowd risk before it happens, and tells you what to do — and why. Any volunteer can report an incident in any language and get an instant, triaged response.

Safety thresholds and crowd numbers are computed by a deterministic engine. Generative AI reasons on top of those facts. It never invents them.

languages triaged
30+
AI briefing latency
< 2s
tests green
483

Any language in · one triaged decision out — every phrase below resolves to SEV-1 · Medical Response

Spanishhay una persona desmayadaBengaliএকজন অজ্ঞান হয়ে পড়েছেHindiएक व्यक्ति बेहोश हो गयाFrenchune personne s'est évanouieArabicشخص فقد وعيهPortugueseuma pessoa desmaiouGermaneine Person ist ohnmächtigJapanese気を失った人がいますSpanishhay una persona desmayadaBengaliএকজন অজ্ঞান হয়ে পড়েছেHindiएक व्यक्ति बेहोश हो गयाFrenchune personne s'est évanouieArabicشخص فقد وعيهPortugueseuma pessoa desmaiouGermaneine Person ist ohnmächtigJapanese気を失った人がいます

The problem

A World Cup control room drowns in feeds.

In 2026, millions of fans pour through turnstiles across 16 host venues. In each control room, a handful of staff watch dozens of camera feeds, density sensors, gate counters and radio channels at once — and a dangerous crowd surge can build in the ninety seconds it takes to notice one screen among many.

The failure mode isn't a lack of data. It's a lack of attention: too many signals, too little time to turn them into a decision anyone can act on and defend.

What NEXUS does

  • Reads every feed continuously and computes a single, honest overall risk level.
  • Rewrites the situation into plain language the moment it changes — no dashboards to decode.
  • Proposes the next action with a real, engine-computed impact number attached.
  • Turns any volunteer’s report, in any language, into a triaged, routed response.

The one idea that makes it trustworthy

The AI has no shape in which to express a safety number.

Generative AI is the product's brain — it writes the briefings, explains the decisions, and understands incident reports in 30+ languages. But it is structurally prevented from deciding anything safety-critical.

The AI proposes

a category

The triage model's output schema has no severity field and no team field. There is no shape in which it can express a triage decision — only the language, the English translation, and one of six categories.

{ language, englishText, category } // no severity. by design.

The engine decides

the severity

A deterministic keyword table scans both the raw report and the translation, so a mistranslation can't launder an emergency into a routine note. A life-safety term overrides the model's category outright.

classifySeverity(report) → SEV-1 · Medical
The consequence is testable, and tested: with Gemini switched off entirely, “hay una persona desmayada” is still SEV-1, still routed to Medical Response. Read the methodology →

How it works

Four layers, one strict direction of trust.

01

Feeds stream in

A seeded simulator models 8 zones and 6 gates of a World Cup venue — densities, gate flow, arrivals — as a pure function of the scenario and elapsed time.

02

The engine computes truth

Pure, tested functions turn raw feeds into density bands, gate utilization, ETAs and one overall risk level. Every safety number originates here — never the LLM.

03

The AI reasons on top

Gemini is handed those facts and asked to explain them: a plain-language briefing, ranked recommendations, and multilingual incident understanding.

04

You get a decision

A briefing that rewrites itself, a recommendation with a computed “110% → 92%” impact, and a triaged incident — each labelled AI or rule-based so you always know the source.

Facts only ever flow engine → AI. The AI is always last, and always optional — if it's slow or unreachable, every panel falls back to a deterministic rule-based answer.

Capabilities

Everything a control room needs, nothing it doesn’t.

Live situational briefings

The overall picture rewritten in plain language the instant it changes — the AI narrates, the engine supplies the numbers.

Decisions with computed impact

Every recommendation carries a real, first-order impact figure from a conservation-of-people flow model. “Open the overflow lane: 110% → 92%.”

Multilingual incident triage

A volunteer reports in any of 30+ languages; NEXUS detects, translates, and triages — with the exact rule that fired quoted for audit.

Resilient by design

AI down? Rate-limited? Every route degrades gracefully to a rule-based answer. The control room never goes dark.

Sustainability at a glance

Energy, waste, water and transit metrics with an AI insight line — the operational and the responsible, side by side.

Accessible & audited

WCAG-AA contrast, full keyboard paths, reduced-motion honoured, and a11y asserted in the test suite — not an afterthought.

The 20-second demo

Watch the room react in real time.

  1. 1

    Open the command center. A normal matchday is filling; overall status reads Normal.

  2. 2

    Pick the “Gate C surge” scenario. The engine reacts and the status pill climbs Normal → High → Critical.

  3. 3

    The AI briefing rewrites itself: “Gate C is taking 215 arrivals a minute but can only process 195…”

  4. 4

    A recommendation appears with reasoning and a computed impact: “Open the overflow lane. 110% → 92%.”

  5. 5

    Switch to Incidents, type in Spanish, and watch it triage SEV-1 → Medical — with the rule quoted.

30+

languages understood

483

automated tests, all green

95%+

safety-engine coverage

100%

deterministic safety math

Under the hood

Engineered to be believed, not just demoed.

Next.js 15React 19TypeScript · strictVertex AI · Gemini 2.5-flashZod-validated boundariesDeterministic engineFirestoreCloud Run

Step into the control room.

See the status climb, watch the AI explain itself, and report an incident in any language — the whole demo runs in under a minute.