Inspiration
Many farmers, especially in rural areas, cannot rely on fast or continuous internet access. At the same time, most AI crop tools depend on cloud services and often stop after giving a single disease prediction.
We wanted to build something more practical: an AI tool that works directly on the farmer’s phone, respects privacy, verifies uncertain results, and helps farmers monitor a plant over time instead of giving only a one-time label.
What it does
AgrioCrop AI is an offline crop intelligence app for plant-disease screening and monitoring.
Farmers can take a photo or upload one from their gallery, and the image is analyzed locally using a YOLOv11x model covering 116 plant-disease classes.
Its main feature, FieldCheck, analyzes three different leaf photos from the same plant and compares the predictions before presenting a result.
AgrioCrop also provides:
Visual affected-area severity estimates Confidence and uncertainty handling My Field plant tracking Follow-up scans Improving, worsening, or stable progress monitoring Scan and FieldCheck history Explainable AI results Safe local action plans Offline Qwen3-1.7B agricultural assistant English and Urdu support Camera and gallery input
AgrioCrop is designed as a screening and monitoring tool, not as a replacement for professional agricultural diagnosis.
How we built it
We built AgrioCrop AI using Flutter and Dart for the mobile application.
The plant-disease pipeline uses a local YOLOv11x model exported to ONNX, allowing inference to run directly on the Android device.
The workflow is:
Camera/Gallery → Image Quality Check → YOLOv11x → Crop Validation → Confidence Checks → Severity Estimation → Local Guidance
For FieldCheck, three images are analyzed independently and a consensus layer compares their predictions.
Plant observations are stored locally so farmers can perform follow-up scans and compare changes over time.
For conversational support, we integrated Qwen3-1.7B Q4_K_M with a local llama.cpp-based runtime. The assistant receives structured scan context such as crop, possible disease, confidence, symptoms, severity, and local guidance.
Firebase is used for authentication and selected cloud features, while the core AI screening and local assistant can work offline.
Challenges we ran into
The biggest challenge was running meaningful AI models directly on a mobile phone.
We had to optimize model loading, memory usage, ONNX preprocessing, inference speed, and local LLM performance.
Another challenge was dealing with uncertainty. A single model prediction should not automatically be treated as a confirmed disease, so we introduced crop filtering, confidence thresholds, multi-photo FieldCheck verification, and clear uncertainty messaging.
We also learned that raw AI predictions are not enough. Turning model output into severity tracking, follow-up monitoring, explainability, and useful farmer actions required additional product logic beyond the model itself.
Accomplishments that we're proud of
We are proud that AgrioCrop AI goes beyond basic plant-disease classification.
The app provides a complete workflow:
Detect → Verify → Explain → Track → Re-scan → Compare → Guide
FieldCheck is especially important because it does not blindly trust a single image. It checks multiple pieces of visual evidence before giving the farmer a stronger conclusion.
We are also proud of the My Field system, which allows farmers to track individual plants and see whether their visual condition appears to improve or worsen over time.
Most importantly, the core AI experience can continue working without a constant internet connection.
What we learned
We learned that building responsible AI requires much more than selecting a powerful model.
Mobile AI must consider device limitations, uncertainty, privacy, user trust, and what happens after a prediction is made.
We also learned that local LLMs become much more useful when they are grounded with structured information from the application instead of being asked to answer everything from memory.
This project reinforced the importance of designing AI systems that know when they are uncertain and help the user collect better evidence instead of pretending to always know the answer.
What's next for AgrioCrop AI
Next, we want to test AgrioCrop across more real-world crops, lighting conditions, phones, and disease stages.
We plan to improve FieldCheck consensus, severity estimation, Urdu agricultural guidance, and mobile inference performance.
We also want to expand the reviewed offline disease knowledge base and develop privacy-preserving regional crop-health signals that could help identify emerging disease patterns without uploading farmers’ plant images or precise farm locations.
Our long-term goal is to make AgrioCrop AI a practical offline crop intelligence companion for farmers anywhere.
Log in or sign up for Devpost to join the conversation.