Module 2 · VOGraph AI + Voyage Graph LLM

Every voyage adds an edge.

A traveller is not a vector, not a tree, not a row in a database. A traveller is a graph. VOGraph AI joins the demand-side Guest Graph to the supply-side Destination Graph through match edges — and the set of all those edges is the matrix ZnVO reasons over.

432 dims

Traveller DNA across 6 categories.

≈97%

Verified feasibility with Z3 SMT + OR-Tools CP-SAT.

0.6%

Feasibility of raw-LLM itinerary output, unverified.

P95 < 2.4s

Five-stage discovery pipeline including intent.

The four-DNA matrix

One engine, four products

Because the match is bidirectionally symmetric, the same computation answers four different questions: what should this traveller do, who should this destination attract, which bundles can this curator win, and where does this supplier truly fit.

Four-DNA matrix diagram with Traveller, Destination, Curator and Supplier DNA feeding a scored bundle match — Score(t,d,c,s) = w_fit·f_TD + w_craft·f_TCD + w_rel·f_SD + w_feas·f_feas · f_feas ∈ {0,1}

Traveller DNA

fit

432 dims · 6 categories · Demand

Who a traveller actually is — rhythm, tolerance, memory, company, constraint. A traveller is not a row in a database; a traveller is a graph.

Destination DNA

craft

432 dims · 9 categories · Supply

Destinations as first-class graph entities with their own attributes, relationships and mutable offering parameters.

Curator DNA

reliability

384 dims · 8 categories × 7 curator types · Trust

The craft of the human who assembles the journey — what they know, where they operate, who trusts them and what they deliver on time.

Supplier DNA

feasibility

384 dims · 8 categories × 5 supplier types · Reliability

Deliverability as data: connectivity depth, on-time performance, cancellation behaviour, claims acceptance.

Feasibility gate

A bundle that cannot be delivered never gets ranked

Language models are excellent at proposing journeys and poor at guaranteeing them. VOGraph passes every candidate bundle through constraint solvers before it can surface, so an itinerary is either verified or it is not shown.

Intent decomposition

Claude Sonnet 4 turns a human ask into constraints in under 200ms.

Graph traversal

Neo4j walks curated, supplied, substitute and evocative relationships.

Constraint proof

Z3 SMT and OR-Tools CP-SAT prove time, geography and capacity feasibility.

Explainable ranking

Each bundle carries the reason it scored, not just the score.

Consent-aware data

A four-level consent model (E/S/D/O) governs what may inform a match.

Tier gates

Airline OTP ≥80% Elite, ≥90% Prestige; IOSA mandatory Elite+; claims acceptance ≥95%.

Under the hood

The stack that carries it

Intent decomposition: Claude Sonnet 4 · ≤200ms
Graph backbone: Neo4j — CURATED_BY, SUPPLIED_BY, SUBSTITUTES_FOR, EVOKES
Learning: PyTorch Geometric GNN · LangGraph · Voyage Graph LLM
Feasibility solvers: Z3 SMT + OR-Tools CP-SAT
Vectors & data: Pinecone 500M+ · MongoDB Atlas · Kafka + Flink · Redis/Feast
Surface: FastAPI · GraphQL · gRPC · Kong · OAuth 2.0 + PKCE · MCP
Runtime: Kubernetes + Istio · Spark + Airflow
Discovery pipeline: 5 stages + intent ≈ 1,680ms · P95 < 2.4s

Graph vocabulary

Edge types

CURATED_BYSUPPLIED_BYSIMILAR_TOSUBSTITUTES_FORCOMPLEMENTSEVOKESCONTAINS

A Pinecone index of 500M+ vectors, MongoDB Atlas documents and a Kafka + Flink stream keep the graph current, while Spark and Airflow rebuild the heavier features.

ZnVO is pre-launch. Nothing on this site implies live volume, signed curators or paying suppliers. Market and regulatory figures are carried from their original sources and should be re-verified before external use.