How We Turned a Citrus Disease Model Into a Working MVP
A real EDVM use case: deploying a research-trained model as a budget-conscious web demo for PhytopathologIA.
Some projects do not need a production platform on day one. They need a credible, working demonstration that proves the idea, supports fundraising, and helps a team move forward with confidence.
That was the situation with PhytopathologIA.
They had already trained a machine-learning model to detect disease in citrus leaves. What they needed from us was the operational layer around that model: a way to upload an image, run inference, and show a result through a simple web interface that non-technical stakeholders could understand immediately.
The brief
PhytopathologIA needed an MVP for investor and stakeholder demos, not a full production deployment.
The target was clear:
- accept a photo of a citrus leaf,
- run the trained model,
- return a prediction,
- make the whole flow accessible in a browser.
From a product perspective, this was about reducing friction between research and demonstration. A promising model in a notebook is valuable, but a live URL changes the conversation.
The constraints
This project had real-world constraints, which is exactly where engineering discipline matters.
- Budget was limited for both development and infrastructure.
- Ongoing hosting costs were not desirable at this stage.
- The team needed something they could show quickly while preparing for the next funding step.
In other words, the right solution was not the biggest solution. It was the smallest complete system that could prove the concept end to end.
What we built
Working closely with PhytopathologIA founders Rodrigo Machado and Sofia Bengoa Luoni, we assembled a lean delivery stack:
- Model deployment for their trained PyTorch classifier.
- A lightweight API layer to receive image uploads and return predictions.
- A simple web interface with drag-and-drop upload for leaf photos.
- A demo-ready runtime using Google Colab and PyNgrok to expose the application through a public URL.
The result was a system that people could actually test, not just talk about.
Why the deployment choice mattered
For this phase, paying for a permanent environment would have created unnecessary cost.
Using Google Colab as the execution environment and PyNgrok as the public entry point gave the team a pragmatic path:
- no recurring hosting bill,
- enough uptime for scheduled demos,
- a fast route from model to usable product.
This is a good example of a principle we use often at EDVM: match the architecture to the business moment. Early validation should be fast, reliable, and cost-aware.
Delivered outcome
| Milestone | Outcome |
|---|---|
| PyTorch model deployment | Inference flow ready to process citrus leaf images |
| Lightweight API | Clear request/response layer for the frontend |
| Web upload interface | Non-technical users could test the classifier in a browser |
| Google Colab + PyNgrok setup | Public demo access without ongoing infrastructure cost |
| MVP scope under tight constraints | A complete prototype shipped for real conversations with investors |
Why this use case matters
This was not a theoretical exercise. It was a real operating artifact for a team trying to validate a product direction.
For PhytopathologIA, the MVP created a bridge between research output and business communication. Instead of presenting only metrics, screenshots, or model notes, they could present a workflow:
- upload a leaf image,
- run the model,
- see the prediction.
That kind of clarity is valuable in any early-stage technical project.
It also speaks to a broader opportunity in agriculture. Computer vision and machine learning are becoming more practical tools for early plant-disease detection, helping teams intervene before disease spreads.
What this says about how we work
At EDVM, we care about building software that matches the problem instead of inflating the scope.
Sometimes that means enterprise architecture. Sometimes it means a lean MVP that proves the core idea without wasting budget. In this case, the right answer was a focused system that turned a trained model into a live demo quickly.
That is the kind of work we want more of: practical engineering, real constraints, and software that moves a business forward.
Want to build something similar?
If you are validating a product, modernizing operations, or shipping a technical MVP, we can help.