I grew up around farms in the Kenya highlands region. Of course, I am a farm boy 😂, and I watched farmers lose entire harvests because they couldn't identify a disease until it was too late. By the time they reached an expert, the damage was done.
Most plant disease apps scan leaves and miss the essential parts of the plant, e.g., the branch, roots, and entire leaf area. This is what I can think of current solutions(With the least research I've done, of course)
Root rot starts underground, Stem borers tunnel through stalks, bark cankers spread silently. By the time symptoms reach the leaves, the farmer is already losing the war.
So I built Shamba-MedCare—"Shamba" (farm) + "Dawa" (medicine) in Swahili, a simple solution focusing on helping farmers, scientists, etc. Checkout here Shamba-MedCare App
The Three Approaches I Considered
Option 1: Train a Custom CNN
The PlantVillage dataset has 50,000+ labeled images, and MobileNetV3-small can hit 99.5% accuracy at just 1MB.
The catch? The images must have perfect lighting and clean backgrounds. My accuracy tanked the moment I tested with real field photos—muddy roots, partial shadows, multiple issues on one plant.
Option 2: Use a Pre-built API (Plantix, PlantVillage Nuru)
These are some of the existing solutions that work on different use cases. They give a classification with a confidence score, Classification alone doesn't save crops.
Option 3: Multimodal LLM (Claude Vision)
This is where things got interesting.
Claude doesn't just classify, it reasons. I can ask it:
"Analyze this tomato leaf. The farmer says spots appeared 2 weeks ago after heavy rain. They can only afford traditional remedies. What's wrong and what should they do?"
And it actually incorporates that context.
The Trade-off I Made
| Feature | Custom CNN | Claude Vision |
|---|---|---|
| Works Offline | Yes | No |
| Contextual Explanations | No | Yes |
| Novel Disease Handling | No | Yes |
| Per-Request Cost | Free | ~$0.01-0.05 |
I chose Claude, not because it's perfect, but because a detailed explanation that is helpful to the farmer and saves a crop is worth more than a fast classification that misses context.
What I Built
The core flow:
Every diagnosis includes:
- Health score (0-100)
- Disease identification with confidence
- Visual heatmap showing WHERE damage is
- Treatment tiers: FREE → Low → Medium → High cost
- Traditional remedies that farmers already trust
Inclusion
I almost built this for all the farmers to be generic. Then I remembered, a 55-year-old farmer in rural Kenya with basic literacy is not the same as a 25-year-old agronomist with a smartphone addiction.
So I added:
- Voice mode: Results read aloud in clear speech while the user is scrolling through
- Huge touch targets: 44px minimum, because field work means rough hands and busy farmers
- Bottom navigation: One-handed operation while holding a plant
- Icon-first design: Pictures over text, and pictures with text
Tech that farmers can't use is just tech for my portfolio
Next in this series: How I structured the prompts to get consistent, budget-aware diagnoses (and damn, the 3 times it completely failed)
Top comments (0)