How I Implemented AI for Plant Diagnosis, Tested It on Other People's Problems, and Why Code Quality is Managed by a Cat with a Witcher Name

Where Did This Idea Come From?

[Here I'll fantasize and make things up, because it sounds cooler :P]
 

Ever since I announced I was working on AI plant disease recognition, messages started pouring in. From friends, from app users, from random gardeners on the internet. The content usually similar:

"Hey, my [plant X] has [something suspicious], do you know what it could be?"

Classic human reflex — when someone builds a tool for diagnosing something, they instantly become a free consultation service in that field.

Instead of turning everyone away, I decided to be smarter. I fed each such report to the AI and checked: is the diagnosis sensible? Is the description in Polish and does it sound like a gardener's answer, not like an encyclopedia spat out from a database? Has the system even gone crazy at all?

Side effect: the app's journal got filled with a dozen entries with plants I don't have, diseases I haven't seen, treatments I've never performed. Cauliflowers with spots. Peas with growth problems. An almond tree that looked like it was having a rough year.

If someone saw this and thought: "this guy has a weird garden" — now you know.

[Now I'm not making things up and getting serious... :P]

Who Was I Helping?

A few cases that passed through my hands (and AI) during testing:

[Since I don't have flowers, in my app money became cauliflower, thuja - my wife's favorite tree - became radish, etc...]

Someone had a cauliflower with yellowing outer leaves and brown spots near the veins. AI: boron deficiency, possibly chlorosis due to alkaline soil pH. Recommendation: check pH, feed with microelements. Sensible.

Someone else had peas that germinated beautifully, then stopped. AI pointed to moisture problems in the substrate during germination or planting too deep. Also logical.

There was also a case of an almond tree with gray-brown growths on branches — AI correctly identified it as probably bacterial canker and proposed a treatment protocol.

All entries landed in the journal as functional tests. Because that's what software testing looks like when you don't have a QA department and you're alone: you use real problems of real people and check if the system gives sensible answers.

By the Way — Ciri

There's one more collaborator I need to mention.

My cat Ciri — yes, exactly, the name from the Witcher saga, because this kitty apparently has some Silesian roots and likes when dramatic words fall — has been monitoring the code quality of Zielna Manufaktura for months.

The method is simple: she sits next to the keyboard, looks at me with clear dissatisfaction when something doesn't work, and physically gets between me and the monitor every time I try to push to production something that isn't ready yet.

Effectiveness at 94%. The remaining 6% are moments when she's sleeping and can't stick her claws into the process.

The AI plant disease recognition module went through her hands — I mean through her paws — several times. The fact that it works is half my merit, and half the consequence of Ciri sitting on the keyboard precisely when I was sending the version with fixes to the server.

I think.

How Does It Work Now and How to Test It?

AI plant disease recognition is available in the app. You take a photo of a sick plant, upload it, get a diagnosis in Polish — what it is, why, and what to do about it.

Each query uses external services that cost real money, so it's not "scan without limit forever". But:

New account: after registration you get several free recognitions to check if it even works on your plants.

Already have an account? Write to me through the form on the website. I'll add some free scans for you — because the more different plants and problems go through the system, the better I can assess whether it works correctly.

Form: 
https://zielnamanufaktura.pl/kontakt 

Community with test entries: 
https://zielnamanufaktura.pl/community

Marek — programmer, gardening enthusiast from Bieszczady and owner of a cat who knows more about code quality than most CI/CD tools

Ciri — Head of QA, unofficially, without employment contract, paid in wet canned food