One photo. AI tells you what's wrong with your strawberry.

楊先生的草莓園 — wide shot — Mr. Yang's strawberry rows · Xinshe, Taichung · betel nut palms in the back, a signature of central Taiwan

Key numbers

2 hr → <10 sec: Inspection time
1 photo: per diagnosis
AI: detects type & location

TL;DR

Shi Jian Strawberry is an organic farm in Xinshe, Taichung. The farmer used to spend two hours a day walking the rows looking for pests and diseases — relying on experience alone with no reference data. Now he opens his phone, takes one photo, and within 10 seconds AI marks the disease type and location. Photos and data are collected and archived automatically.

On the farmer's phone

One photo. AI boxes the disease in under 10 seconds.

Mr. Yang only sees one screen: open the PWA, take a photo, AI boxes the disease and labels it — with reference data and treatment recommendations. No app install, no typing, no separate system.

Interactive demo · click to start · audio toggle inside

PWA disease detection — capture, detect, diagnose, log — PWA — capture, detect, diagnose, log

What NEXUPIRA does here

Mr. Yang does not need to know what YOLO or a Workflow Engine is. He just knows that a photo gives him an answer. Behind the scenes, the full NEXUPIRA stack is running. When the spectrometer comes online, it is one more step in the same Workflow — no rewrite, no new platform.

Vision module

YOLO object detection, identifying disease type and location.
Event Bus

Detection results trigger events; anomalies notify automatically.
Workflow Engine

Photo → notification → data archival, one DAG runs it all.
Google Cloud

Everything runs on the cloud; the farmer operates from his phone.

Under the hood

How the AI learned to spot disease.

Behind '<10-second detection' is a traditional ML pipeline: collecting disease samples from the field, labeling class and location in Roboflow one by one, training a YOLO model until mAP converges to ~90%, then deploying to the PWA. Actual workflow, below.

Roboflow labeling interface — 01 — Roboflow: class label + bounding box

Model training metrics — 02 — Training metrics: mAP converging, Box / Class / Object loss dropping

Phases

✓
Live

PWA photo capture identifies strawberry and leaf diseases in under 10 seconds.
◎
In progress

Partnering with NCIR to add spectral analysis — measure sweetness and pesticide residue without picking the fruit.

Spectrum, illustrated