Krishi AI

Inspiration

The motivation to develop this agentic AI model for farmers arose from witnessing how intensely farmers battle crop diseases and the absence of timely advice. Most rural farmers lack easy access to agricultural specialists, and before they notice a disease, it already results in significant losses in yield. With how platforms such as WhatsApp and Meta have simplified the way people communicate, We wished to make something similar but for agriculture—a smart assistant that is convenient to use for farmers. The plan was to integrate AI with affordable technology in a way that farmers would only need to send a photo of their crops, respond to a few queries, and receive precise information regarding potential diseases and cures instantly. This project came into being based on the dream of minimizing the loss of crops, equipping farmers with information, and ultimately supporting food security and sustainable agriculture.

What it does

Krishi AI is an agentic AI assistant designed to support farmers in protecting and improving their crops. By simply sending a photo of their plants on WhatsApp, farmers can instantly get insights about possible diseases along with practical solutions. The system is enhanced with weather and finance APIs, so it can give recommendations that are more accurate and suited to local conditions. Our goal with Krishi AI is to make advanced agricultural support accessible, affordable, and easy to use for every farmer.

How we built it

We created a multi-modal AI by fine-tuning a MobileNetV2 computer vision model for visual disease detection, achieving high accuracy on the PlantVillage dataset. This diagnostic tool is integrated with a 7B parameter LLM running locally via llama.cpp with CUDA acceleration, which provides farmers with detailed, multilingual conversational advice and remedies based on the visual diagnosis. After training the model we tired integrating it with a WhatsApp Business Account using the Facebook API, and combined weather and finance APIs for better insights.

Challenges we ran into

Key challenges included debugging the image data pipeline for the vision model and managing GPU memory to run the 7B LLM efficiently offline using GGUF quantization. A significant effort was made to ensure the AI's output, both visual and text-based, was simple, actionable, and farmer-friendly rather than overly technical. We also faced the challenge of not being able to deploy it on a server due to financial constraints, which limited the scalability of our solution.

Accomplishments that we're proud of

We successfully built an end-to-end system that achieves 97.04% validation accuracy in visual disease identification and deployed a 7B parameter LLM to run locally and offline without cloud costs. We are proud of creating a tool that can guide a farmer from a simple photo to a complete, conversational treatment plan.

What we learned

This project demonstrated the power of combining specialized models. We learned that transfer learning is key for high-accuracy vision tasks, while quantization is essential for running large-scale language models efficiently on local hardware. Above all, we learned the importance of a user-centric design in making complex AI useful. We also learned how to connect AI models with real-world platforms using APIs and design impactful solutions while balancing technical and financial constraints.

What's next for Krishi AI

Our next step is to improve the model's real-world performance by training it on a more diverse, crowd-sourced dataset of images. The ultimate goal is to deploy this multi-modal system into a highly accessible mobile app or chatbot that can serve as a comprehensive AI advisor for farmers globally.