Consistent tobacco leaf grading on the floor you already run
Grading decides what a grower is paid and what a merchant can sell, yet on a busy receiving floor the call is made in seconds, by eye, by a classer who has already handled thousands of bales that day. This deployment puts a camera-based second opinion beside the classer: as each bale is scanned it is photographed, a model proposes a grade, and the person confirms or overrides it. Every bale then carries a picture and a stated reason, so a disagreement three weeks later is settled from the record rather than from memory.
The operational challenge on a receiving floor
Grading is a judgement about leaf position, colour, body, ripeness and damage, and two experienced classers can honestly disagree on the same bale. In the marketing season the volume arriving in one morning far exceeds the trained classers available, and consistency drifts as the shift wears on. That drift surfaces as grower disputes at payment, as rework when a buyer rejects a lot after shipment, and as arguments nobody can settle because nobody photographed the bale. Non-tobacco related material, meaning twine, plastic and similar contamination, is usually found by the customer rather than at intake.
The infrastructure the deployment uses
Almost everything needed was already on site: CCTV over the receiving line, a recorder on the local network, weighbridge and platform scales, a bale ticket printer with barcode scanning, and a grower and contract management system. We added an edge computer, meaning a rugged PC that runs the vision model on site rather than in a distant data centre, beside the recorder. Lighting over two grading positions was corrected and a colour reference card fitted in frame. No receiving-line cameras were replaced and the ticketing workflow was left intact.
What was built
This is a grading assistant, not an automatic grader. Scanning a bale ticket triggers image capture from the fixed camera positions; the edge computer classifies the images against the client's own grade schedule and returns a suggested grade with a confidence level and a short reason. The classer sees this on a console at the table and either accepts it with one press or overrides it, picking a reason from a short list.
How it runs day to day
A delivery is weighed and each bale scanned as it reaches the table, so the suggestion appears before the classer has finished handling the leaf and acts as a prompt rather than a delay. Suspected contamination raises a separate alert, which the classer confirms or dismisses. Supervisors work from a daily view of override rates by classer and by grade, which is where genuine drift, an unusual crop or a knocked camera shows up early.
Integration points
Bale records are matched to grower and contract using the existing ticket barcode, so grading data lands against the right delivery with nothing rekeyed. Final grades and weights are pushed to the stock and settlement systems that produce grower payments, and images stay attached to the bale for the agreed dispute window. Grower notifications of accepted deliveries run through InOne CRM over WhatsApp, and season and regulatory reporting is exported from the same register.
What it covers
Suggested grade with confidence
Every scanned bale gets a proposed grade against your own grade schedule, with a confidence level and a stated reason. Low-confidence bales are flagged as uncertain rather than dressed up as certain.
Photographic bale register
Images, grades, timestamps and classer identity are stored against the bale number, so a grower query or a customer claim is answered from the record.
Contamination alerts at intake
Twine, plastic and string visible on a bale raise an alert at the table. Catching it at intake costs far less than a rejection after shipment.
Override capture and drift review
Overrides are recorded with a reason, showing supervisors where classers and the model disagree and giving us the data the model is retrained on.
Runs when the link is down
Classification runs on the edge computer on site, so grading continues through internet outages and records sync afterwards.
Consistent capture conditions
Fixed viewpoints, corrected lighting and an in-frame colour reference keep captures comparable across a season. Colour judgement is only as good as the light it is made under.
How it works
Site assessment
We walk the receiving floor during a working shift, document the cameras, scanners, scales, ticketing and grower system in place, and agree the grade schedule the system must speak.
Integration and observation
The edge computer, console and camera positions go in and connect to existing ticketing and grower records. The system then observes only: it suggests while classers grade as they always have, and we compare the two without influencing anyone.
Rules and guardrails
From that observation and your historical graded bales we set the confidence threshold, override reasons, alert behaviour and escalation for disputed bales. Nothing goes live until the floor manager has signed off on how the system behaves when it is unsure.
Go live and optimise
The console goes live at the tables and we review weekly, tuning against confirmed overrides. A first system is typically four to six weeks from assessment to go live.
Questions buyers ask
Does this replace our classers?
No. It gives each classer a second opinion and a permanent record, and the person keeps the final word on every bale. The system exists to hold that expertise steady across a long shift.
How accurate is it?
We do not publish an accuracy figure, because a number from someone else's crop and floor tells you nothing useful. It is measured on your own bales during the observe-only period, against the grades your classers assigned, and reported before anyone decides to go live.
Do we have to buy new cameras?
Usually not on the receiving line. Grading positions are the exception, as colour judgement needs stable lighting and a fixed viewpoint, so the assessment states which positions need a lighting change or a dedicated camera first.
Will it work with our grade codes?
Yes. It is configured against the grade schedule you already use with your buyers and growers, and can be reconfigured between seasons.
What happens in a power or internet outage?
Grading carries on. The model runs on site and records queue locally until the link returns, and the edge computer shares the protected power already serving the recorder.
Who can see the images and how long are they kept?
Data lives in your environment or a dedicated tenant, access is controlled and logged, and retention is agreed in writing before go-live, normally covering the season's dispute and claim window.
Book a platform briefing
A 15-minute technical discussion about your receiving floor: what your cameras, scanners and grower system already support, and what an observe-only period would need to prove before you commit. Contact sales@eigenstatesystems.com or WhatsApp +263 77 636 6999.