Skip to content

Attractor Engine

Unstructured data in.Structured intelligence out.

NK Technologies builds and licenses the Attractor Engine. It turns emails, PDFs and scans into validated, structured data, and runs the agents that act on it for hours or days at a time.

Dubai · DIFC

Get in touch →
NKengineEmail triagerouted to ops deskExtractionshipment.jsonClassificationcustoms declarationValidation2 fields to reviewSummariesdaily exceptions briefAgent runsrun_8f2c waiting

The engine

A processing engine, not a document tool.

One engine turns any unstructured input into structured, validated output. Every use case runs on the same runtime — capabilities are configuration, not new code. Document Intelligence, email triage, workflow automation, and agents all run on the same engine.

  1. 01Input arrives, and context is assembled.
  2. 02The model runs, and the output passes deterministic validation.
  3. 03A confidence score gates it — above the threshold, auto-accepted; below, routed to a person.
  4. 04Every correction feeds back into the deployment and improves the next run.

The differentiator is the gate, not the model call. Output you can put into a production system, because everything below a measured confidence threshold goes to a human — and every correction makes the next run better.

The gate

Nothing ships unchecked.

Between the model and your systems sits a gate. It is what makes the output safe to put into production.

  1. 01

    Deterministic validation

    Every output is checked against fixed rules first — structure, types, required fields. This part is not the model; it passes or it does not.

  2. 02

    An engineered confidence score

    Each result carries an engineered confidence score for how much it can be trusted.

  3. 03

    The threshold

    Above the threshold, output is accepted automatically. Below it, it is routed to a person before anything is used.

  4. 04

    Corrections compound

    Every human correction folds back in within the deployment, so the same work needs the gate less over time. Nothing is pooled across customers.

Long-running agentic workflows

Agents that keep working after the first answer.

Real logistics work doesn't finish in one call. A run waits for a carrier, pauses for a person, retries a portal that's down, and picks up exactly where it stopped, whether that's minutes or days later.

run_8f2c · Demurrage dispute, masked container number

Run completed in 2 days, 4 hours, 48 minutes.

  1. Invoice received

    Demurrage charge parsed from the carrier's PDF.

    Mon 08:14

  2. Evidence gathered

    Gate-out, free-time terms and port delay notices matched.

    Mon 08:16

  3. Dispute drafted

    Claim drafted with the timeline attached.

    Mon 08:21

  4. Claim approved

    An ops lead checked the claim and approved sending it.

    Mon 09:05

  5. Waiting on carrier

    Claim sent. The run is parked with its state saved. No polling loop.

    paused 2d 2h

  6. Credit accepted

    Carrier offered partial credit. An ops lead accepted it.

    Wed 11:40

  7. Credit booked

    Credit note posted to the WMS and the case closed.

    Wed 13:02

  • Durable state

    Every step is checkpointed. A restart or deploy resumes the run; it never starts over.

  • People in the loop

    Runs pause for an approval and continue as soon as someone signs off.

  • Waits on the world

    Emails, portal updates and timers wake a parked run. Nothing sits polling.

  • Full audit trail

    Each decision, document and model call is recorded against the run.

The Runs surface — illustrative

Packs

The engine is generic. Capabilities ship as configuration.

  • Document Intelligence

    Compound files in; segmented and extracted structured records out. A mixed document is split into its parts, the fields that matter are pulled and validated, and anything uncertain is routed to a person before it is accepted.

  • Email Triage

    Classifies, contextualises, and routes inbound mail, holding drafted replies behind the same confidence gate until they clear the threshold or a person signs off.

  • Workflow Automation

    Runs event-triggered processes as graphs on the same runtime, routing high-impact steps to human approval before they proceed.

  • Agents

    Plans and executes multi-step, tool-using work under the same deterministic validation, confidence scoring, and human gates as every other output — nothing acts unchecked.

Logistics use cases

Built on freight, warehouse and customs paperwork.

These are the jobs it handles across freight, warehouse and customs work.

  • Booking intake

    Booking requests arrive by email in any layout. They become structured orders in the WMS without retyping.

    emailorder

  • Bills of lading

    Shipper, consignee, container numbers and seals extracted and checked against the booking.

    pdfshipment

  • Customs pre-check

    HS codes, values and origin compared across the invoice and packing list before filing.

    invoice + packing listdeclaration draft

  • Proof of delivery

    Phone photos of signed delivery notes read, matched to the stop and flagged when anything's missing.

    scanpod record

  • Freight invoice audit

    Carrier charges checked line by line against contracted rates. Overbilling goes to review.

    xlsxaudit result

  • ETA exceptions

    Delay notices read as they arrive. Affected orders are found and customer notices drafted for sign-off.

    carrier noticenotice drafts

On the warehouse floor

One inbound truck, six decisions the engine makes.

A truck docks. The engine reads the paperwork, matches it to the order and holds what doesn't add up for a person. Then it hands storing and picking to the warehouse robots. The engine decides; the robots execute.

Illustrative floor, documents and figures; not customer data or results.

  1. Delivery note read. 14 lines · confidence 0.97.
  2. Matched to order PO-7731. 13 of 14 lines agree.
  3. Line 6: 12 cartons, 10 ordered. Held for a person.
  4. Slot C-04 assigned · task to robot fleet. 13 matched lines · AMR-03 stores the pallet.
  5. Pick list · task to robot fleet. AMR-07 picks 3 incl. C-04 · then dock D4.
  6. Carrier ETA +3h on D5. 2 outbound orders re-planned.
  7. Result: 14 lines read, 93% auto-accepted, 156 m of route saved, 1 line held for a person.

Across the network

From the vessel to the warehouse door, ahead of time.

One container vessel crosses from Shanghai to Rotterdam. Before it berths, the engine has read the manifest, caught a delay, pre-checked customs and booked the dock slots in the warehouses that receive the cargo.

Illustrative network and figures; not customer locations or results.

  1. Manifest read at departure. 2,140 containers · 186 bills of lading.
  2. Transshipment matched. 312 containers → Jebel Ali feeder.
  3. Suez queue: ETA +36h. Rotterdam & Hamburg plans moved.
  4. HS code conflict on 1 container. held for a customs specialist.
  5. Dock slots booked before berthing. 14 trucks · 2 warehouses.
  6. Hamburg receiving plan ready. labour & racks planned for Thursday.
  7. Result: 2,140 containers checked, 2 issues caught early, 22 dock slots booked across 2 warehouses.

What goes in, what comes out

Messy input. Clean, validated output.

The engine takes mixed, unstructured input and returns structured records you can put straight into a system.

In

  • Compound documents
  • Inbound mail
  • Event streams
  • Mixed, unstructured sources

Out

  • Segmented, extracted, validated records
  • Routed decisions
  • An audit trail of what was accepted and what a person reviewed

Deployment

Two ways it ships.

Headless

Behind your systems.

The engine runs behind your existing systems. Output lands where your work already happens.

Workbench

A surface of its own.

The engine plus its own operational surface — review queue, audit trail, document history, exports. Your team works in it directly.

The Workbench surface — illustrative

Sectors

Where the engine is applied.

  • Maritime & supply chain

    Structured, validated output from the unstructured input that moves cargo — documents, messages, and data.

  • Logistics & warehouse operations

    Operational signal pulled and validated from mixed, unstructured inbound.

  • Back-office document & communication workflows

    High-volume back-office work across documents, email, and messages — gated and audited.

Structure

Built in-house. Operated by our delivery partner.

  • NK Technologies

    Developer & licensor

    Develops and licenses the engine.

  • Attractor Labs

    Delivery partner & operator (DACH)

    Co-founded by NK Technologies.

  • Shipsoul

    Technical partnership

    Maritime logistics.

How we work

Built in-house. Maintained as products.

Built in-house

Architecture, engineering, and AI systems are built in-house, not assembled from parts.

Production standards

The engine is in production, held to production standards — not demo ones.

Accountable past handover

The engine is maintained by the team that built it. Responsibility does not end at the licence.

Trust & data handling

Isolated, in-region, and yours.

Isolated per customer
Deployed per customer in isolated infrastructure. No shared runtime, no pooled data.
In-region data residency
Data and processing stay in the region the customer chooses.
GDPR posture
The customer is the controller; Attractor Labs GmbH is the processor.
You own your data
You own your documents, results, and corrections, and can export them at any time.

FAQ

Questions, answered.

Is this a product or a service?
NK Technologies builds and maintains the engine. Attractor Labs operates it for customers. It ships headless behind existing systems, or as the Workbench. Delivery partners run it in their markets.
What happens when the engine is not sure?
It does not guess. Output below the confidence threshold is routed to a person before anything is accepted.
Does it handle more than documents?
Yes. Document Intelligence is one pack; email triage, workflow automation, and agents run on the same engine.
What about GDPR?
Attractor Labs GmbH is the processor; the customer is the controller.

Contact

Direct line, no forms.

For partnerships, product, or anything else.

info@nktech.ai