Camera that catches the defect that slips through

The problem

The sizer measures the avocado; it does not always see what is on its skin.

Natural fruit variation in size, ripeness and texture -the quality-control pain the plant declared- means the automatic size sorter doesn't always catch subtle external defects at real line speed. That fruit reaches the distributor: a claim, a full box return, lost trust in an export market where reputation is decided lot by lot.

• Avocado varies naturally in size, ripeness and skin texture, and the automatic sizer does not always tell a subtle defect from normal skin at real line speed.
• Bruises, skin spots, early anthracnose, black tip and over-ripe fruit slip past visual sampling at the end of the line, because at that speed a person sees a sample, not the fruit.
• That fruit keeps ripening on the way to market, and the defect shows up at the distributor as a claim or a full box returned, often with the whole lot put under suspicion.
• In export, reputation is decided lot by lot: a returned box costs far more than the fruit inside it, because the freight, the claim handling and the customer's doubt about the next lot all come with it.

How it fits the IRIS system

Edge — every fruit classified in milliseconds, on a local GPU, before the box.

iLEAN Edge: an industrial camera over the packing line, with a local Edge GPU and a CNN trained on good/bad examples of the specific fruit (bruises, spots, black tip, over-ripeness), infers in milliseconds and rejects the defective piece before final packing.

A defect on an avocado grows after packing: a faint bruise at the plant is a dark patch at the destination. Catching it on the line is the last moment it costs one fruit instead of one box. And the rejected fruit still has a second life in the pulper.

See the full IRIS architecture →

Before and after

Sampling at the end of the line versus Edge on every fruit

Aspect	Today	With iLEAN Edge
Share of fruit inspected	A sample, when there is time	All of it
Subtle bruise or early anthracnose	Passes at line speed	Classified and rejected
Point where the bruise is found	At the distributor	Before final packing
What the model knows	Generic sizing rules	Your own good and bad fruit
Where inference runs	—	A GPU at the line, in milliseconds
Borderline fruit	Packed anyway	Escalated to a person

Partial human visual sampling → 100% fruit inspection. Visual-defect claims → drastic reduction.

Impact estimate

Impact estimate — to be validated with your numbers.

The block below is an estimate to be validated against your plant's actual data. We put it forward so the committee has an order of magnitude; we refine it during the assessment.

• Estimated payback 5-12 months, depending on your current reject ratio and on how many of your lots go to export markets that claim for visual defects.
• The return comes from fewer export claims and fewer returned boxes, and from the fruit that is diverted to pulp or oil instead of being written off at the destination.
• Inspection goes from partial visual sampling to every fruit on the line, at the line's real speed.
• And claims for visual defects drop sharply, which protects the lots you have not shipped yet and the distributor relationship behind them.

Estimated payback 5-12 months depending on current reject ratio, from fewer export claims and returns. *Estimate to be validated.*

And the fair question from the production manager

“Every avocado looks different — won't it reject good fruit and cut into the packout?” — natural variation is exactly why the model is trained on your own fruit, varieties and lighting, not on a generic catalog. Classifying a known list of defects under a fixed camera is an anchored task, where the best models drop below 1.5% error [1], and borderline fruit is escalated to the line supervisor instead of being rejected blindly. Every decision the supervisor makes on those borderline fruits goes back into training, so the model learns your own tolerance.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.