SPROUT

SPROUT

Smart Plant Recovery and Optimization Utility Tool

Take the walkthrough Live demo

Photograph a sick leaf. SPROUT names the likely disease, asks a few yes or no questions written from 237 university guides, and tells you when it can’t tell.

Tomato leaves with brown ringed spots and yellowing edges

Tomato with early blight, the kind of photo SPROUT starts from.

How a check works

  1. The odds after the photo: potato late blight 78 percent, tomato late blight 21 percent, and the first question, is this a potato or a tomato plant

    The photo gives the odds

    A ConvNeXt V2 vision model scores 29 conditions across 14 plants in under half a second. When the top suspects belong to different plants, it asks which plant first.

  2. Question 2: are there irregularly shaped, wet-looking patches on the leaf, with NC State Extension named as the source

    Questions come from the guides

    While you read, a background worker searches 237 university extension guides and writes yes or no questions from the symptom passages. Code keeps the one that best separates the suspects.

  3. The final answer, tomato late blight 97 percent sure, with care advice and the UMass Extension guide it came from

    A diagnosis you can check

    Each answer updates the odds in code, so the same answers always give the same result. Once one disease clears 93%, you get it with care advice and the guide it came from.

  4. The chat: can this plant be saved, answered with three numbered sources from UMass, NC State and Clemson Extension

    Then ask anything

    The chat remembers the conversation and answers from the same guides with numbered sources. Describe a symptom and the odds move again.

The right disease is in the top three from the photo alone, before a single question
96%
From photo to diagnosis, median on a laptop CPU
0.4 s
From a tap to the next question, median over 56 taps
6.5 ms
University extension guides it reads from
237

What’s under the hood

A vision model trained on a GPU, a search engine over the guides, and a small server that keeps every number in code.

See the architecture
  • Vision modelConvNeXt V2 Tiny fine-tuned in PyTorch on a CUDA GPU with mixed precision, on 19,260 lab and field photos, exported to ONNX for CPU serving.
  • Guide searchHybrid keyword and embedding search over 237 guides, with a cross-encoder reranker on top.
  • Question engineOdds updated by code, next question chosen by expected information gain, questions and advice written ahead in background threads.
  • Experiment trackingEvery training run and result logged in MLflow, data versioned with DVC, every published number traced to a run record.

Aakash Senthilnathan

Statistics and Data Science at UC Santa Barbara. I built SPROUT end to end: the data pipeline, the model, the search, the server and this site.