Skip to content
Planning & productionMining & Raw MaterialsWaste & Recycling

Sorting purity from camera images and analyses

demi links camera images and analyses with throughput, plant status and deliveries and shows why the purity of a fraction drops.

Today

How pure a fraction is decides its price, whether PE film, paper and board, wood or metal. Yet purity is measured only in spot samples, often only at the buyer, and a drop is explained by gut feeling: too much throughput, a dirty sensor, a bad delivery. Images from the belt cameras, sorting analyses, plant data and weighbridge tickets sit side by side, but nobody puts them on one timeline.

An example

At Wertora Recycling, the PE film fraction from the commercial waste sorting plant now reaches only the second quality grade at the buyer, whose analyses show 88 instead of 94 % purity. The agent links the images from the belt camera behind the near-infrared sorter with the plant's own sorting analyses, the throughput at the pre-shredder and the deliveries of the last four weeks: purity drops as soon as throughput rises above 18 t/h and films overlap on the belt, worst with containers from two construction customers carrying wet paper that sticks to the film. It proposes throttling to 15 t/h for such deliveries. The plant manager tests this for a week, and the buyer's next analysis shows 93 %.
Fictional company, typical workflow.

Your data, wherever it lives

The data on how pure your fractions are is spread out today: images in the cameras above belts and sorters, sorting analyses in spreadsheets, throughput and faults in the plant control, tickets at the weighbridge, buyers' findings in e-mails and the ERP. demi reads, understands and links them per line and hour, with your rights. The AI sees during the shift when purity drops and why, and your plant management decides what changes.

Where the data lives today

  • Photos, video & audio
  • Office files
  • Machines & sensors
  • E-mails
  • ERP
Data LayerOne view of everything, with your rights

How demi solves it

Three layers in this order: connect the data first, then the rights, then the AI.

  1. 1 · ConnectData Layer

    Data sources in this case: images from the cameras above the belts and at the sorters, sorting analyses of input and fractions in spreadsheets, throughput, faults and settings from the plant control, deliveries and weighbridge tickets, sales and buyers' quality findings in the ERP. demi replaces nothing and puts it all on one timeline per line and fraction. For every hour you see what came in, how the plant ran and how pure it came out.

  2. 2 · PermitPermission Layer

    Plant management, shift leads and quality see images and results; only sales and management see buyer prices and supplier data. A changed setting or feedback to a waste supplier is approved by plant management; every approval is recorded.

  3. 3 · ActIntelligence Layer

    The agent continuously estimates the composition of input and fractions from the belt images and checks the estimate against every sorting analysis. It shows when purity falls and compares those periods with throughput, belt loading, faults, settings and deliveries. It backs up the factor that explains a drop with the hours and images behind it; the language model only explains the result.

How it runs

  1. 1

    demi continuously reads belt images, sorting analyses, plant data and weighbridge tickets.

  2. 2

    The agent estimates the composition of input and fractions and checks it against every analysis.

  3. 3

    When a fraction's purity falls, it alerts the shift, not only after the buyer's analysis.

  4. 4

    It shows whether throughput, belt loading, sensor, setting or delivery is behind it, with images as evidence.

  5. 5

    Plant management decides on settings, throttling or feedback to the waste supplier.

What changes

  • Purer fractions and better prices from buyers

  • Drops caught during the shift, not weeks later

  • Evidence per delivery for talks with customers

What does it bring you?

Check it with your own numbers

Every starting value is an assumption you can change. The calculation sits right below, step by step.

Assumption, editable

Your numbers

Result

+90 hfreed per year

Assumptions and calculation

demi links images, analyses and plant data and shows why purity drops; you decide what changes.

60 min → 42 min per case

How does it run today?

What is already digital saves less: the factor lowers the share demi brings.

How we calculate

25 purity reviews × 60 min × 30% × 1.0 (manual) × 12 months = 5,400 min = 90 h

Value per year, before demi costs

Freed per year+90 h

Assumption: labor cost of $48.62 per hour in US manufacturing in June 2026 (BLS), plus workplace, IT and overhead.

= Value per year, before demi costs

+$5,850

gained per year

90 h ÷ 1,760 h per role = 0.1 full-time roles

An estimate, not a promise. The starting values are assumptions for an industrial company with 150 to 500 people.

Combine with more cases and get the package

Questions about this use case

Do we need new cameras to measure purity?

Not necessarily. Many sorters and belts already have cameras or sensors whose images and counts can be used, and photos from sorting analyses exist anyway. Where a view of a fraction is missing, the agent shows which camera would close the gap. How the images get to demi, we clarify in conversation.

How accurate is purity from images?

A camera only sees the surface of the stream, so the agent checks every estimate against your sorting analyses and the buyers' findings and shows the deviation openly. For decisions, the trend over hours counts, not the single image. It does not replace lab analyses, but it shows you when one is due.

Can demi see which delivery pulls purity down?

Yes, through the timeline: weighbridge ticket, tipping time and bunker level show which loads are running through the plant at that moment. If drops pile up with certain suppliers, the agent prepares evidence with images. Sales and plant management decide whether and how to talk to the customer.

Does demi change anything on the sorting plant itself?

No. The agent proposes throughput, settings or a sensor cleaning and shows what the proposal rests on. It is put into practice only once plant management or the shift approves it.

See this case with your data

In the demo we show you the workflow on your own systems.

Your configurator

No use cases yet

Your configurator is still empty

Add a use case from any card or case page and see what it brings per year, with the calculation. Free, no sign-up, no commitment: try it out and change any number.

Browse use cases