Computer Vision Engineer — Crop Disease Detection

Chloe Martin

Freelance · Mid-level · Nairobi

This role is no longer accepting applications

Similar open roles you can apply to now:

About the role

We build precision agriculture software for smallholder and mid-size farms in East Africa and South Asia. Our mobile application is used by 140,000 farmers across Kenya, Tanzania, Uganda, and India to track crop health, access agronomic advice, and connect to input supply and output market services. The computer vision component of our app — which allows farmers to photograph a crop and receive a disease identification and treatment recommendation — was built two years ago using a third-party model. It works for the six disease classes it was trained on. It does not work for the additional fourteen disease classes our agronomists have identified as high-priority for our user base, and it performs poorly on the image quality typical of low-end Android cameras used in field conditions. We are commissioning a 10-week project to build a replacement model trained on our own annotated dataset. We have 47,000 annotated crop images across 20 disease classes, collected from field surveys over the past three years. If the prototype meets accuracy thresholds, we will integrate it into the mobile app and discuss an ongoing engineering relationship for maintenance and new disease class expansion.

Key Deliverables

List the expected deliverables for this project.

Audit the existing annotated dataset for quality and class balance, and produce a written assessment with recommendations
Train and evaluate crop disease classification models across 20 disease classes to agreed accuracy thresholds
Develop image preprocessing and augmentation strategies appropriate to field photography conditions
Optimise inference performance for mobile deployment on mid-range Android devices
Deliver a documented model package and evaluation report suitable for our mobile engineering team to integrate

Requirements

Technical stack needed for this mission.

4+ years of computer vision engineering with experience training classification or detection models on real-world, non-benchmark datasets
PyTorch for training and evaluating image classification models on small, carefully curated domain-specific datasets
Python for data pipeline development, augmentation strategies, and model evaluation
Experience with data augmentation techniques for low-quality, variable-lighting imagery — field agriculture datasets are not clean
Scikit-learn for baseline model analysis and evaluation metric calculation across multi-class disease classification
Docker for reproducible training and evaluation environments the integration team can build on after the engagement
Familiarity with model quantisation or optimisation for mobile deployment is a significant advantage