AgroGina

The Problem

Moroccan agricultural cooperatives needed a way to monitor crop health across hundreds of distributed farms, many in rural areas with unreliable internet. Traditional scouting methods were slow, subjective, and expensive. AgroGina’s vision: satellite-driven agronomy insights accessible to smallholder farmers, combined with ERP tools for inventory, sales, and workforce management.

Key challenges:

• Satellite data integration — Process multi-spectral imagery from Sentinel-2 to compute NDVI (vegetation health index)
• Geospatial complexity — Manage thousands of farm parcels with polygon boundaries, crop rotation history, and yield forecasts
• Multi-tenancy — Each cooperative sees only their farms, teams, and agronomic data (strict org_id isolation)
• Offline-first UX — Field agents need to record observations and interventions without connectivity, syncing later
• Dual domain architecture — Business logic (ERP) and satellite processing (Google Earth Engine) require different tech stacks

Our Approach

We split the platform into two specialized backends with a unified React frontend:

Business API (NestJS + PostgreSQL)

• Multi-tenant ERP — Every query filtered by org_id at the database level (Postgres RLS policies)
• Farm management — CRUD for parcels, crops, planting schedules, harvest records
• Team & roles — Org-scoped user accounts with agronomist, field-agent, and admin roles
• Inventory & sales — Track fertilizer stock, equipment, and crop sales per cooperative
• REST + GraphQL — NestJS provides both interfaces; frontend uses GraphQL for complex nested queries

Satellite Processing API (Python FastAPI + Google Earth Engine)

• GEE integration — Server-side Earth Engine Python SDK for NDVI computation, time-series analysis
• Batch processing — Celery workers fetch new Sentinel-2 imagery weekly, compute vegetation indices per parcel
• Geospatial queries — PostGIS extensions for polygon intersection, area calculations, buffer zones
• Caching layer — Redis caches NDVI time-series to avoid redundant GEE API calls (quota optimization)

Frontend (React SPA)

• Leaflet maps — Interactive farm boundary editing, NDVI heatmap overlays
• Offline sync — IndexedDB caches farm data; service worker queues mutations for later sync
• TanStack Query — Optimistic updates with retry logic for flaky rural connectivity
• PWA support — Installable on Android tablets used by field agents

Multi-Tenancy Architecture

Every database table includes org_id (UUID). Postgres Row-Level Security (RLS) policies enforce:

CREATE POLICY tenant_isolation ON farms
  USING (org_id = current_setting('app.current_org_id')::uuid);

NestJS sets the session variable on every request via a global interceptor. This guarantees zero cross-tenant leakage even if application code has bugs.

The Stack

• Frontend: React 18, TanStack Router + Query, Leaflet + React-Leaflet, IndexedDB, Service Worker
• Business Backend: NestJS, TypeORM, PostgreSQL 15 with PostGIS, GraphQL (Apollo)
• Satellite Backend: Python FastAPI, Google Earth Engine Python API, Celery + Redis, GeoPandas
• Database: PostgreSQL with PostGIS extension, RLS policies, TimescaleDB for time-series NDVI
• Infrastructure: Docker Compose, nginx reverse proxy, DigitalOcean Droplets, automated backups

Outcome

AgroGina launched in September 2025 and now serves 12 cooperatives managing 4,200 hectares across Morocco:

• 8,500+ NDVI analyses — Automated weekly vegetation health reports for every enrolled parcel
• Offline-first proven — Field agents in rural Chefchaouen and Ouarzazate record data during week-long trips, syncing when back in town
• Zero tenant data leaks — RLS policies passed penetration testing; no cross-org access incidents
• 40% faster anomaly detection — Satellite alerts flag irrigation issues or pest infestations 2-3 weeks earlier than manual scouting

Engineering Lessons

Dual backends > monolith for this domain: Keeping NestJS (business rules, fast CRUD) separate from FastAPI (CPU-heavy GEE processing) let us scale and deploy them independently. The GEE API is slow (300-2000ms per query), so caching is mandatory.

RLS is the right tool for strict multi-tenancy: Row-Level Security moved tenant isolation from application code (error-prone) to the database (enforced). We sleep better knowing a forgotten .where('org_id', ...) clause won’t leak data.

Offline-first is non-negotiable for rural Morocco: Early beta users abandoned the app when they couldn’t record field observations offline. Adding IndexedDB + service worker sync turned NPS from 6 to 9.

NDVI alone isn’t enough: Farmers requested prescriptive recommendations (“apply 20kg/ha nitrogen”) based on NDVI trends. We added a rules engine that interprets satellite data through agronomic models, which became the platform’s killer feature.

Need multi-tenant SaaS with complex domain logic? Let’s talk.