CropSentry

Technology

A real AI stack, built for the field.

CropSentry combines on-device computer vision, cloud-scale training pipelines, and GPU-accelerated inference. Here is exactly how it works.

The AI pipeline

From dataset to diagnosis.

1. Data

12,000+ labeled leaf images across maize, cassava, tomato, cocoa — collected in partnership with Nigerian cooperatives. Goal: 250k images by end of Phase 2.

2. Models

Fine-tuned EfficientNet-B4 (classification) and DINOv2 (feature extraction), distilled into a 12 MB on-device model using PyTorch Mobile and ONNX Runtime.

3. Training

NVIDIA A100 GPUs (via AWS EC2 P4d) with CUDA + cuDNN. Mixed-precision training, automated hyperparameter sweeps, weekly retraining as new field data arrives.

4. Inference

Two-tier inference: on-device for sub-100ms offline scans, cloud fallback (Triton Inference Server + TensorRT) for complex cases.

5. Outbreak intelligence

Federated scan data (anonymized) feeds a regional disease-spread model on Amazon Bedrock — predicting hotspots before farmers report symptoms.

6. Safety

Every diagnosis carries a confidence score. Below 70%, we route the scan to a human agronomist via our extension-officer dashboard.

AWS

How we use AWS.

AWS is our cloud backbone — from secure storage of farmer data to multimodal advice generation in local languages.

  • • S3 — versioned storage for raw training images, model artifacts, and scan history.
  • • RDS (PostgreSQL) — farmer profiles, scan records, farm geometries, outbreak events.
  • • Lambda + API Gateway — serverless inference API for low-cost scaling.
  • • ECS Fargate — long-running Triton Inference Server instances behind a load balancer.
  • • Amazon Bedrock — Claude/Llama models generate localized treatment advice in Yoruba, Hausa, Igbo, Swahili.
  • • SageMaker — managed training jobs and model registry.
  • • CloudFront + S3 — fast app delivery across African regions.
  • • SNS + Pinpoint — outbreak SMS alerts to farmers without smartphones.
  • • CloudWatch + X-Ray — observability across the inference path.

NVIDIA

How we use NVIDIA.

NVIDIA accelerates every step where speed matters — training, inference, and edge deployment.

  • • CUDA + cuDNN — core GPU acceleration for PyTorch training jobs on A100s.
  • • TensorRT — INT8 quantization and graph optimization drops inference latency from 240ms to under 80ms.
  • • Triton Inference Server — multi-model serving so vision and language models share GPU memory efficiently.
  • • NVIDIA NeMo — fine-tuning of language models for African-language advice generation.
  • • RAPIDS — GPU-accelerated data pipelines for image preprocessing at scale.
  • • Jetson Nano / Orin — planned edge deployment for extension-officer field kits and farm gateways.
  • • NVIDIA Inception — applied; aligning with NVIDIA's startup AI program for technical mentorship.