Consortium AI
Sign in
Confidence-gated · human-in-the-loop · consortium-powered

Catch the trade-finance fraud your bank can't see alone

Consortium AI screens LLM-extracted invoices against your confidence threshold, routes doubt to your reviewers, and shares anonymized risk signals across member banks — so the same fraudulent supplier can't hit five institutions unnoticed.

The demo runs the real gate in your browser — no signup. Free tier, no card required.

How it works

One pipeline, four checkpoints

01Live

Extract

Invoice metadata enters through a strict schema — pasted JSON on every plan, AI extraction from raw documents on Pro, both behind the same validation.

02Live

Gate

Every critical field must clear your bank's confidence threshold. A missing score counts as zero and fails. A high overall score can never mask a failing field.

03Live

Review

Anything below threshold queues for a registered, active reviewer. Decisions are attributed to an authenticated identity and recorded immutably.

04Live

Share

Approved transactions are anonymized — identities become SHA-256 hashes — and shared as risk signals other member banks can match against.

The network is the moat

One bank sees an invoice. The consortium sees a pattern.

When two member banks process transactions naming the same supplier, buyer, or invoice, their nodes match on hashes — and each side sees the overlap without either exposing a customer. Amount, currency, and date stay clear for risk scoring; identities stay hashed end to end. Every bank that joins makes every other member harder to defraud.

Two deployments wire together point-to-point with a shared inbox key — no central data lake, no third party holding your customers.

Supplier overlapseen by 1 other node

8191bace4a5683b9…035ccca2

node-demo-partner500,000 USDissued 2026-06-30
An actual match card from the Signals page — identities hashed, risk in clear.

Built for compliance teams

Guarantees, not promises

These are properties of the architecture, enforced in code and verifiable in the audit trail — not lines from a sales deck.

Low confidence can never auto-approve

The gate is a pure function with fail-safe rules: missing confidence scores as 0.0 and fails, threshold comparisons never relax on bad config, and every failing field is reported.

Identities never leave in clear

Supplier, buyer, and invoice identifiers cross the node boundary only as SHA-256 hashes. The consortium inbox rejects raw identity fields at the API — it is enforced, not promised.

Everything is on the record

The audit trail is append-only and written exclusively by the server. Gating decisions, reviewer actions, threshold changes — regulators see the whole chain, unedited.

AI document extraction (Pro)

Hand the model a raw invoice and get structured, confidence-scored fields back — behind prompt-injection guards, hard input caps, and the same fail-safe gate as manual input. The model can never grant itself auto-approval.

Try it

Start free. Scale to the consortium.

Screen transactions on the free tier today. Upgrade for AI extraction, and talk to us about bank-node deployments and point-to-point consortium wiring when your institution is ready.