AI & Models Overview

Nexa's AI footprint is small and deliberate. Every model has a narrow job, a deterministic fallback, and a human in the loop for anything irreversible. Every AI feature is powered by Nexa Trained Models on Google — there is no third-party LLM API for the customer to provision, contract, or operate.

The four models

#	Model	Type	What it does
1	Flight-disruption predictor	Nexa Trained Model on Google	Predicts cancel probability + expected delay for upcoming flights from weather / ops / traffic features
2	Allocation scorer	In-process, deterministic	Ranks hotel candidates by cost + distance + consolidation
3	Policy synthesizer	Nexa Trained Model on Google	Converts natural-language policy descriptions into structured policies
4	Exception agent	Nexa Trained Model on Google	Triages manual-review items; proposes actions to human operators

Models 1, 3, and 4 are Nexa Trained Models running on Google infrastructure; model 2 is a deterministic scorer that we've called a "model" deliberately because it can be tuned per policy.

Flight-disruption predictor

A dual-head Nexa Trained Model trained on weather pressure, labor pressure, traffic pressure, hazard pressure, flight-ops pressure, destination pressure, recent cancel rate, and seasonality. It returns:

{
  "cancelProbability": 0.0721,
  "predictedDelayMinutes": 23,
  "confidenceScore": 0.78
}

A deterministic baseline runs as a fallback when the trained model is unreachable — Nexa never blocks on model availability.

Used to pre-warm contingencies: Nexa can begin hotel search and demand planning before the airline confirms a disruption, shaving minutes off the response curve.

→ Deep dives: Flight Predictor — API guide · Disruption Dashboard · Inside the Flight Predictor

Allocation scorer

Deterministic, versioned, and tunable per policy. The scorer ranks hotel candidates via a weighted combination of:

• Cost (dominant)
• Distance from the airport
• Consolidation — a bonus for putting more passengers in the same hotel

Policy constraints (stars, amenities, price cap) act as gates before scoring. The scorer is explainable: the allocation wave persists every input and output so you can answer "why this hotel?" during a finance review.

→ Deep dive: Allocation

Policy synthesizer

A structured-output task. The operator writes:

For MAD business and first class, 4+ star only, within 15 km, must have 24-hour reception and airport shuttle. No budget chains.

and the synthesizer returns a structured policy draft with provenance (source: "ai", confidence, prompt) that a supervisor reviews and activates. No activation is automatic.

→ Deep dive: AI policy synthesis

Exception agent

A tool-using agent with a small allow-list of read-only tools. Its job is to look at a manual-review item and propose up to three actions the operator could take, each with an estimated cost and a reason.

The agent:

• Reads the case, the allocation wave, and the provider errors.
• Uses searchAlternateHotels, computePolicyRelaxationCost, and classifyFailureCause tools.
• Emits structured recommendations.

It does not book, cancel, or notify — those tools are not in its allow-list. Every recommendation is a human-approved action.

→ Deep dive: Exception agent

Design principles

• Trained where determinism buys you nothing. Flight disruption prediction is a pattern-recognition problem — the right tool is a trained model.
• Deterministic where explainability matters. Allocation is a finance-visible decision; a weighted scorer beats an opaque model for audit.
• Trained models for language. Policy synthesis and exception triage are language-shaped problems where AI adds leverage — but always with a human check.
• Fallbacks are first-class. Every AI call has a deterministic fallback path; Nexa will never fail a booking because a model is down.
• Allow-list over prompt engineering. Agents' capabilities are constrained by the tools available, not by the prompt. A leaky prompt can't make a read-only tool write.

Disabling AI

Every AI capability is a per-tenant toggle in the operations console (under Settings → AI):

• Exception agent — on / off
• Policy synthesizer — on / off
• Flight predictor — on / off

Nexa runs end-to-end with all AI disabled. You lose the nice-to-haves (pre-warmed contingencies, one-click policy drafts, agent-recommended rebooks) but the core pipeline — policy → case → demand → allocation → booking → notification — is fully deterministic.