How Agririgo reads a field photograph

This is the walkthrough we would give you standing over your shoulder. It matters because a tool that tells you your crop is stressed without telling you how it decided is a tool you cannot argue with, and an agronomist who cannot argue with a number should not act on it.

The five stages

1. 1

   The picture arrives, and stays

   You drop a drone orthomosaic or a phone photograph into Rigo Studio. It is read into your browser's memory. It is not uploaded. If you close the tab now, nothing of it exists anywhere but your own machine.

2. 2

   The vegetation overlay is computed on your device

   Studio computes a vegetation index across the image, in your browser, using your own processor. This is arithmetic on pixel values, not a model: the same picture gives the same overlay every time, on any machine, and we can show you the formula. That is what lets it run offline and instantly.

3. 3

   The image is checked before a model ever sees it

   Not every photograph can support every claim. A frame at five millimetres per pixel cannot show you a lesion on a leaf, so Studio refuses to run the crop disease classifier on it rather than returning a confident answer it has no basis for. Declining to answer is a feature.

4. 4

   Rigo AI reads what the overlay cannot say

   An overlay shows you where the crop is under stress. It does not tell you whether that is water, nitrogen or disease. That is the question Rigo AI answers, against open crop disease classifiers, and it is the one part of the pipeline that runs on our servers rather than your device.

5. 5

   The finding is ranked, not narrated

   Severity and priority are computed from the findings by formula. They are deliberately not left to the model, because a model asked to rank its own output will do it fluently and inconsistently. An early version let the model set severity and it called a healthy field 'high', which the interface then painted red.

Why everything carries a confidence

Every finding arrives with a number attached, and that number enters the record exactly once. Combining confidences twice is how systems end up certain about something no single measurement supported. If Studio is not sure, the screen says it is not sure, and you get to weigh that against what you can see with your own eyes in the field.

A finding you can check is worth more than a finding that sounds certain.

What happens over a season

One photograph is an observation. The value shows up on the second one. Every observation is tied to a field and a date, so Studio compares this week against last month and this season against last, and tells you what changed in numbers rather than asking you to remember. Trends are fitted with methods that tolerate the odd bad reading, because field data always contains one.

All of this works without an account. Rigo AI needs to know who is asking, because it runs on our hardware, but the overlay and the labelling do not.