Now in public preview

Durable AI Reconciliation for Logistics Document Sets

Serverless TypeScript agent that extracts, cross-checks, and prepares resolution actions for shipment discrepancies through durable Inngest orchestration.

Choose an example Both use sample documents

The Live Orchestration Engine

Follow the fan-in extraction and discrepancy pipeline as each document becomes structured data, then a shipment-level reconciliation result.

Ready to start Choose an example, then run it to follow the document check.

0.0s Recorded example

Live Workflow

The highlighted stage shows the current durable step.

Stage 0

Receive documents

Place both files under the same shipment and confirm they are ready to read.

> POST /demo/runs
> emit document.uploaded

Stage 1

Extraction

Find parties, quantities, weights, dates, and item references using typed schemas.

> AI SDK generateObject
> persist extraction_run

Stage 2

Reconciliation

Fan in the document set, claim a set hash, and compare shared shipment facts.

> wait for N docs
> deterministic compare

Stage 3

Agent proposal

Clear the shipment or prepare a human-approved action for the discrepancy.

> propose action
> wait for approval

Shipment Documents

0 of 2 ready

Commercial invoice INV-DEMO-200 - Widget A - 1,200 units

Waiting

Packing list PL-DEMO-200 - Widget A - 1,150 units

Waiting

Active Discrepancies

Waiting for both documents

No discrepancy yet Once both documents are extracted, Reconciler compares the fields that must agree and prepares a review path only if one is needed.

Item quantities

Parties and addresses

Dates and weights

Serverless & Durable

Built on Vercel and Inngest so long-running LLM extraction, fan-in reconciliation, and human approval pauses never depend on one held request.

Automatic retries on transient provider failures
Durable fan-in of related documents before comparison
Content-hash and set-hash idempotency anchors

export const reconcileShipment = inngest.createFunction(
  { id: "reconcile-shipment", debounce: { period: "60s" } },
  { event: "document.extracted" },
  async ({ event, step }) => {
    const required = await step.run("expected-set", loadExpectedSet);
    const ready = await step.run("check-complete", checkComplete);
    if (!ready) return { status: "waiting" };

    const fresh = await step.run("dedupe", claimSetHash);
    if (!fresh) return { status: "already-done" };

    return step.run("compare", reconcile);
  }
);

Vercel Inngest Drizzle ORM AI SDK

Experience the Durable Workflow

Run the sandboxed sample set to watch extraction, reconciliation, and the review proposal update in place.