ElderLeaf

ElderLeaf is an "old hand" for new growers. A beginner takes one photo of their plot and gets what an experienced grower sees at a glance: how many plants there are, how much of the ground the crop covers, how much of the green is weeds, whether that is normal for the plants' age (FAO-56 growth stages), how many litres of water the plot needs today (from local Open-Meteo weather). It keeps a logbook of each plot and compares two dates, with the grower's own notes in between. Everything is explained in plain language, in a tone matched to the grower's experience. Meta's models do the seeing and the explaining. A Muse Spark Open agent looks at the photo and decides what SAM 3.1 should search for: ranked open-vocabulary phrases with fallbacks, plus a scan mode for beds with several crops. SAM 3.1 then draws every plant and weed, so every number is tied to a mask the grower can tap. A second Muse Spark agent writes the report and can ask SAM to look closer at a possible problem. No model output ever becomes a number: code computes all measurements. A validator rejects any sentence that contains a number not found in the measurements, or names a chemical, a dose or a disease. We tested each design choice on 1,010 labelled images from PhenoBench and CropAndWeed. Letting the agent choose SAM's words raised crop F1 from 0.35 to 0.48 across eight crops. On photos with two crops, scan mode reached F1 0.78, where searching by crop names alone reached 0.04. ElderLeaf is live as a web app (elderleaf.vercel.app) and an Android app, running on Modal and Supabase at about $0.013 per photo.