The Problem We Were Solving

India's 150 million farmers lose an estimated ₹50,000–90,000 crore annually to crop diseases that could have been caught earlier. There are roughly 800 plant pathologists for 140 million farming households. By the time a farmer reaches an agronomist, a detectable fungal infection has already spread across multiple plants.

The AgroBot project started with a simple question: can we put accurate crop disease detection in every farmer's hands, on hardware they can afford, running entirely offline in the field?

The Hardware Constraint

₹4,000 is the price point that matters in rural India — within reach of a farmer cooperative's shared equipment budget. It corresponds roughly to a Raspberry Pi 4 with camera module and weatherproof enclosure. Emphatically not a GPU server.

We chose the Raspberry Pi 4 (4GB) as our reference platform — 1.8GHz quad-core ARM Cortex-A72, no dedicated ML accelerator. If we could hit accuracy and latency targets here, the model would also run well on more capable edge hardware like the NVIDIA Jetson Nano.

Model Architecture and Training

The detection model is a MobileNetV3-Small backbone with a custom classification head, quantized to INT8 for edge deployment. Full model footprint: 3.2MB. We trained on 47,000 labeled crop images spanning 12 disease categories across rice, wheat, and cotton.

Critically, the training data was collected in Indian field conditions: variable lighting, soil-dusted leaves, phone-camera quality at various angles. Models trained on clean laboratory images fail badly in actual field deployment — a mistake we designed around from the start.

The 38ms Number

38 milliseconds is the average inference time on the Raspberry Pi 4 for a single 224×224 input image, using TensorFlow Lite with INT8 quantization and XNNPACK acceleration. Fast enough for real-time video at ~26 fps — smooth enough for a drone-mounted camera doing a continuous field scan.

The 94.8% accuracy figure is measured on our held-out test set, which deliberately includes challenging cases: early-stage infections, images from unseen geographic locations, and poor lighting conditions. We use F1 score as our primary metric since disease classes are imbalanced in field conditions.

Voice Advisory in Local Languages

Detecting a disease is only half the problem. Communicating the result and recommended action to a farmer who may not be literate in standard Hindi is the other half. Rama STT and Shiva TTS integrate directly with the AgroBot inference pipeline to provide voice-based advisory in the farmer's local language.

The advisory output — pesticide recommendation, application rate, timing — is generated by Krishna LLM with an agri-domain system prompt. Full pipeline from image capture to voice advisory runs in under 2 seconds. All processing is local: no cloud connectivity required, no data leaves the device.