A wilting leaf can mean a dozen different things — overwatering, underwatering, nitrogen deficiency, low light, root rot, spider mites, or just old age — and a single photo rarely settles which one. Most plant-ID apps guess straight from the image: confidently, and often wrong, because the information that actually resolves the ambiguity — watering habits, drainage, light hours, recent weather — was never in the picture. Plantopia works the way a clinician would instead: it asks before it answers.
Upload a photo and Plantopia identifies the species using two independent methods (Vision model and Pl@ntNet), extracts the visible symptoms, then pauses to ask clarifying questions chosen for your specific case. It returns not a single verdict but a ranked differential — two or three candidate diagnoses, each with the evidence for and against it, plus a five-minute test to tell them apart — followed by a treatment plan that escalates from the least invasive option first. Every plant gets a history; each new diagnosis is checked against the one before it to say whether it's improving, and a chat lets you ask follow-up questions scoped to that plant, sourced and cited.
Under the hood, it's a LangGraph agent with two graphs: an explicit state machine for diagnosis, so identification always precedes diagnosis and the clarifying-question pause always fires, and a ReAct loop for open-ended chat. Rather than trusting similarity search alone, the agent first reasons about which disorders are worth reading about, then retrieves those specific documents from a curated corpus by name. Outdoor diagnoses also pull real weather, matched to the day and location the photo was actually taken rather than the day it was uploaded.
The stack is FastAPI, PostgreSQL with pgvector, React and LangGraph, evaluated against a 28-case golden set at 89.3% top-1 and 96.4% top-3 diagnostic accuracy. Observability is handled with LangSmith and LangGraph Studio. It's also built to be honest about its limits: it says when it can't be sure rather than guessing, strips GPS and other metadata from every upload, and treats anything it retrieves from the web as untrusted data rather than instructions to follow.
Anyone can register with just an email and password — there's no invite list or approval step. New accounts are unverified until the address is confirmed via an emailed link, so nothing owner-scoped (plants, diagnoses, history) is reachable until then. Sign-in keeps you logged in securely without repeated logins.
