Agrosight AI

• 

  Computer Vision performance metrics

• 

• 

  Diagnosis via web app

• 

  Diagnosis via Whatsapp

Agrosight AI – Bringing Field-Tested Computer Vision to African Agriculture

About the Project

Agrosight AI is a WhatsApp and web-based crop disease and pest diagnosis platform built to serve rural farmers in Kenya and, eventually, across Africa. It combines computer vision, AI microservices, and local knowledge to deliver instant, actionable insights in farmers’ hands — without requiring expensive smartphones, fast internet, or technical skills.

The mission:

“Perception as a service — robust, localized, field-tested visual intelligence for African agriculture.”

Inspiration

The idea was born from two observations:

The data gap – There is very little high-quality, locally sourced agricultural image data for African crops and pests.

The last-mile problem – Rural farmers often lack access to timely, affordable, and trustworthy agronomic advice.

By fusing my background in AI with the simplicity of WhatsApp, I saw a way to bridge the gap and bring expert-level diagnosis directly into the farmer’s daily tools.

🛠 How I Built It

1. Data Collection & Annotation

Sourced ~20,000 images across maize, tomato, tea, and rice diseases and pests.

Manually labeled 10,962 images by hand in YOLO format — a monk-like process I likened to copying 5,000 pages of scripture.

Balanced underrepresented classes to 500 samples each using Albumentations for offline augmentation.

1. Model Training

Used YOLOv8 for its speed and field deployability.

Trained on Colab, iterating hyperparameters for optimal precision-recall balance.

Exported best-performing weights for inference.

1. Deployment

Built a Dockerized FastAPI microservice for AI inference.

Integrated with a Django backend for data storage and farmer records.

Created a Next.js web frontend for optional browser-based access.

Developed a WhatsApp onboarding and diagnosis pipeline, allowing farmers to send crop photos and receive instant insights.

📚 What I Learned The real bottleneck in AI isn’t model architecture — it’s data.

Designing for low-resource environments requires ruthless simplicity and reliability.

Farmers respond better to actionable advice than raw probabilities — e.g., "Apply neem oil within 48 hours" "Apply neem oil within 48 hours" instead of "Confidence: 92%" "Confidence: 92%"

⚡ Challenges Faced Data Scarcity – Most open datasets are biased toward non-African crop varieties. I had to curate and annotate my own dataset.

Offline Augmentation – Roboflow failed to balance classes correctly, forcing me to rebuild the augmentation pipeline in Python for full control.

Low-Bandwidth Optimization – Compressing models and API responses for rural areas without losing accuracy.

End-to-End Integration – Ensuring FastAPI, Django, WhatsApp, and the frontend worked seamlessly in production.

🌍 Impact & Future Agrosight AI is now ready for soft piloting with rural farmer groups, schools, and local agricultural organizations. The long-term vision includes:

Expanding dataset coverage to all major African crops.

Embedding AI in affordable sorting, grading, and harvesting hardware.

Building Africa’s first Computer Vision hub for agriculture.

Built With

• albumentations-(data-augmentation)
• cloudinary-(media-storage)-messaging-integration:-whatsapp-business-api-(twilio)
• django-rest-framework-(backend)
• javascript
• languages:-python
• next.js-(frontend)-ai/ml:-yolov8-(crop-disease-&-pest-detection)
• openai-api-(fallback-insights)-databases:-postgresql-(primary)
• pandas
• python-decouple-(config)
• railway/render-(backend-&-ai-engine)-task-queue:-celery-with-redis-broker-(upstash)-authentication:-jwt-(simplejwt)-with-refresh-token-flow-other-tools:-github-actions-(ci/cd)
• typescript-frameworks:-fastapi-(ai-inference-service)
• vercel-(frontend)
• websockets-(real-time-web-updates)-cloud-&-deployment:-docker