CXAiQ pulls every scan, delivery attempt, customs hold, call, chat and driver handoff into one Neo4j knowledge graph — so an express carrier can deflect the contact before it happens, explain exactly why contacts spike, and intercept fraud the rule engines miss. One substrate, three pillars, every decision audited and reversible.
Scans, sort decisions, customs holds, handover handshakes, route deviations — by the time a customer calls, the cost is already locked in, and the connections that explain it are scattered across a dozen authoritative-but-disconnected systems.
The disciplined Amplinth loop, tuned to express logistics — events become a graph, the graph answers honestly, action stays bounded, and every outcome is attributed, not assumed.
LLM extraction maps every call, chat & email to an L1→L2→L3 intent taxonomy on the graph. No black box.
GDS — PageRank, Louvain, WCC, FastRP — and a calibrated, monitored call-probability model write scores back as node properties.
Proactive SMS/email through a consent → frequency → quiet-hours gate, with a one-click kill switch. Freezes are reversible.
An attribution loop earns the avoided-contact count; contact cost avoided becomes an EVA ledger line.
Every action is a citable record on an append-only log — usable in a partner-data audit, a complaint, or a post-mortem.
Three value pillars — Visibility, Deflection and Fraud — plus the composable ingestion that feeds them, all reading one graph that puts the customer and the parcel at the centre.
Intent taxonomy (L1→L2→L3) and root-cause PageRank rank the most impactful preventable causes; Louvain intent communities, drift detection (volume / velocity / community shift), emerging-pattern surfacing, a Sankey interaction→intent→root-cause flow and a UMAP scatter make the cascade legible to Network Ops.
A calibrated call-risk model ranks high-risk cohorts; node-similarity "digital lookalikes" find phone-preferring customers who behave like self-servers; durable customer segments persist across runs; proactive SMS & email notifications close the loop — with an attribution loop so the avoided-contact number is earned, not assumed.
Every send passes a bounded-action eligibility gate: kill switch → live consent (never cached) → frequency cap → quiet hours, each suppression logged with a typed reason for cheap audit. A read-only Cypher validator guards the graph assistant, and an anonymisation discipline keeps raw text and PII out of the graph entirely.
WCC identity rings expose customers sharing addresses and burner emails; reversible, audited account freezes act on them; driver-customer collusion (Louvain) and Isolation-Forest novel-typology scoring over FastRP embeddings surface schemes the rule engines miss; de-identified GraphRAG gives investigators the evidence subgraph, not just a score.
A six-stage adapter pipeline (Connector → Parser → Validator → Mapper → Enricher → Emitter) lands canonical Avro events on Kafka under a versioned Schema Registry; CSV upload-and-apply, streaming graph loaders, a Visual Mapper and a medallion lakehouse mean a new source is a configuration exercise, not a rebuild.
Visibility, Deflection and Fraud aren't three products stitched together: they read the same Neo4j source of truth, share the same GDS machinery (the Louvain that clusters intents also detects collusion), and write scores back transactionally so reads never see a half-applied score. The marginal cost of the third pillar is low because the substrate is shared.
The figures below are the steady-state targets an express carrier sets at the outset — the basis for the EVA business case, not realized client results. CXAiQ is built to attribute progress against the carrier's own baseline, so each one becomes an earned, defensible number.
Emerging themes that take 30+ days to surface today — the target is under 7 days, with drift & emerging-pattern detection already built.
All headline outcome figures are operator-set targets from the value model, not realized client benchmarks. They become realized, attributed figures once a design-partner baseline is in place.
Connect the six-stage adapter framework to TMS/WMS/CRM (CDC), booking webhooks, MQTT telematics, partner SFTP/EDIFACT, M365 and Genesys transcripts. Reason over a Neo4j graph with a GDS registry running on cadence. Act through a bounded, reversible eligibility gate — and Prove every avoided contact on an append-only audit log that ties straight to EVA.
The hardest questions — "which root causes are most impactful?", "which customers share identity?", "which drivers cluster suspiciously?" — are literally PageRank, WCC and Louvain. Solving them on a graph substrate is an order of magnitude less code than the relational equivalent, and dashboards can't claim it.
The graph never sees raw text. An anonymisation gateway, append-only audit as an invariant, a read-only-allowlist assistant, de-identified GraphRAG and a bounded-action eligibility gate with a kill switch make explainability concrete — every action citable, reviewable and reversible.
A deflection attribution loop and a calibrated, monitored model mean the contact-avoidance number is earned — the basis for an EVA ledger rather than a vanity metric. We label every target as a target until a baseline makes it real.
Book a 30-minute demo on your hardest contact-volume problem — we'll walk the Visibility, Deflection and Fraud loop on the graph, and the EVA business case behind it.