Technology
How the platform works today, and the GPU-accelerated geospatial models we are building next.
Sodagro stack
Live todayArchitecture today
Four layers, all in use now.
Responsive, server-rendered web app built with Next.js and React. Works in any phone or desktop browser with nothing to install.
Serverless API routes for chat, disease scans, fields and plans, with authentication, per-plan daily AI quotas and rate limiting.
Frontier multimodal large language models for image diagnosis and agronomy advice, grounded in each field's details and forecast, plus a rule-based agronomic risk engine.
PostgreSQL with row-level security for every table, private object storage for crop photos, and live weather forecast data.
Inside a disease scan
The live Crop Disease Scan pipeline, step by step.
JPEG, PNG or WebP up to 4 MB, checked and saved to private storage.
Crop, sowing date, soil and irrigation from the selected field are attached.
A vision-capable model examines the photo together with that context.
A schema-validated result: crop, diagnosis, severity, confidence, look-alikes and treatments.
Shown to the user and saved to history. If anything fails, the daily quota is refunded.
Weather alerts run on a separate rules engine: each field's 7-day forecast is checked against thresholds for heat stress, frost, heavy rain, strong wind, dry spells (using evapotranspiration) and fungal-disease conditions.
Roadmap · Building next
The next stage moves from general-purpose AI to models trained on Indian agricultural data. These workloads are planned and not live yet.
Ingest multispectral satellite imagery, mask clouds, and compute NDVI and related vegetation-index time series for every registered field.
GPU-accelerated raster processing at scale
Fine-tune vision models on images of Indian crops, varieties and field conditions, collected with farmer consent through pilots.
GPU training and optimised low-latency inference
Time-series models combining weather, crop stage and satellite signals to estimate yield and risk at field level.
GPU training on multi-season datasets
A multimodal model that understands imagery, weather and agronomy across crops, regions and seasons.
Large-scale distributed GPU training
Research & collaboration
Good agricultural AI needs field data and domain expertise. We are looking for partners to help build and validate the next generation of Sodagro models.
Field validation, crop and disease expertise, and evaluation of model outputs.
Remote sensing, computer vision and climate-risk modelling for Indian agriculture.
Responsibly sourced, consented datasets of crop images, field records and outcomes.
Any data used for research is collected with consent and handled under our Privacy Policy.
partnerships@sodagro.comGet started
Create a free account and try the disease scan and AI Advisor on your own crops.