how the ai works
Vision is the product, not a feature.
Naming a real plant from a phone photo and reasoning about care in a specific home was impossible for software a few years ago. Here is exactly how Fernfriend does it, and where it stops.
1. Identify the plant
Your photo is resized in your browser (longest side 1,024 pixels, metadata stripped) and sent to a vision model with a strict instruction set. The model returns a species guess, a scientific name, a confidence between 0 and 1, and up to two genuinely plausible look-alikes. If the photo is not a plant, it says so instead of guessing.
2. Read your home's light
Alongside the photo, the model gets your window direction, distance from the glass, climate, hemisphere, and today's date, which gives the season. It reasons about how much light that exact spot gets and whether it suits the plant, and writes that as a short light note.
3. Tune the rhythm
From the plant and the spot, the model sets a watering range in days, light need, feeding interval, humidity, repotting, and a season note. On Gardener the rhythm keeps shifting with the seasons on our side: winter stretches it, summer tightens it, feeding rests in winter.
4. Read sick leaves
A leaf photo plus the plant's name and spot goes to a separate instruction set modeled on a calm nursery worker. It returns the most likely issue, a category, the visible clues, an honest confidence, 3 to 6 recovery steps in order, what would mean it is getting worse, and other possibilities.
Model-agnostic, on purpose.
Fernfriend does not depend on one company's model. Every reading goes through a thin routing layer that tries vision models from more than one provider in a configured order and takes the first valid answer. When a better or cheaper model ships, we switch with a configuration change, not a rewrite.
Every answer is checked.
Model replies must match a strict schema: confidence between 0 and 1, watering ranges inside sane bounds, a pet-safety status from a fixed list. A reply that does not match is sent back once with the errors; if it still fails, you get a clear message, never a half-formed card.
Where it stops
- It is guidance, not a guarantee. Readings are likely, never certain, and say so.
- Rare cultivars, seedlings, and blurry or partial photos lower confidence. We show that rather than hide it.
- Pet-toxicity flags follow ASPCA lists. If a pet has eaten a plant, call your vet.
- For a precious or valuable plant, confirm with a local nursery before anything drastic.
- We do not train models on your photos. See Your data.