Computer Vision Engineer (AgriTech)

Company Description:   

The Company offers an agri-fintech platform that helps financial institutions and agricultural stakeholders make smarter lending and farm management decisions by turning real-time crop and field data into actionable insights. The platform combines advanced crop monitoring, expert agronomic knowledge, and data analytics to reduce risk, improve productivity, and enhance access to finance in agricultural markets, particularly in emerging regions. 
                                                                                                                            

Project Description:     
We are developing a Computer Vision-based orchard analytics system that segments perennial orchards into tree canopies, crop rows, and in-row spaces using 15 cm RGB satellite imagery (Vantor).

The segmentation outputs are converted into GIS-ready polygons and agronomic metrics, including:

• Row orientation
• Row spacing and length
• Number of rows and canopies
• Canopy cover per row
• Missing-tree indicators

The system will power an agrifintech MVP and is designed to scale globally.

Role Overview:

We are looking for a Computer Vision Specialist / Data Scientist with strong Computer Vision experience to design, train, and deploy a semantic segmentation model for orchard structure detection from high-resolution satellite imagery.

This is a hands-on role focused on:

• Semantic segmentation modeling
• Large-scale geospatial raster processing
• Model experimentation and optimization
• Integration with a GIS analytics pipeline

You will work closely with the Technical Lead (GIS & Remote Sensing), Labeling Specialist, and Backend/DevOps support.

Key Responsibilities:

Model Development:

• Design and implement semantic segmentation models (DeepLabv3+, U-Net or similar).
• Train models on tiled 15 cm RGB satellite imagery (512×512 px tiles).
• Implement augmentation strategies and loss functions for structured agricultural scenes.
• Run hyperparameter tuning and experiment tracking.
• Improve model performance using pseudo-labeling and semi-supervised techniques.

Data & Pipeline Engineering:

• Build preprocessing pipelines for raster tiling and dataset preparation.
• Work with GDAL / Rasterio outputs and georeferenced datasets.
• Implement sliding-window inference over large mosaics.
• Optimize inference performance for large-area batch processing.

Geospatial Integration:

• Support mask post-processing (morphology, skeletonization, polygon extraction).
• Collaborate on the derivation of row centerlines and structural metrics.
• Prepare outputs compatible with GeoTIFF, GeoJSON, and PostGIS.

Evaluation & Validation:

• Define and track segmentation metrics (IoU, precision/recall).
• Perform structured validation at the field and row level.
• Work with agronomic stakeholders to validate meaningful outputs.

Required Qualifications:

• 3+ years of experience in Computer Vision.
• Strong proficiency in PyTorch (preferred) or TensorFlow.
• Hands-on experience with semantic segmentation architectures.
• Experience working with high-resolution imagery (satellite, drone, or aerial).
• Experience with semi-supervised learning.
• Experience with geospatial data. Familiarity with geospatial coordinate systems and projections.

Preferred Qualifications

• Experience with agricultural or remote sensing projects.
• Exposure to transformer-based segmentation models (e.g., SegFormer).
• Experience integrating ML pipelines into backend systems.
• Knowledge of PostGIS or spatial databases.

Technical Environment

• PyTorch + segmentation frameworks
• GDAL / Rasterio
• LabelMe for annotation
• PostGIS for spatial storage
• Cloud GPU infrastructure (RTX 3090 / T4 / A100 class)
• Satellite imagery (15 cm RGB Vantor)

What Success Looks Like (MVP)

• Stable segmentation of orchard row structure across Uzbekistan pilot region.
• Accurate row centerline extraction.
• Reliable row-level agronomic metrics.

Inference pipeline scalable to large areas with minimal manual intervention.    

Working conditions:                                                                                                                                  

Note: Client requires work-time overlap within their time zone (specified below).

Mon – Fri 9-5 (Uzbekistan time)