Skip to content
About Croptimum

What Croptimum is, and how it works.

Croptimum is a digital farming assistant and AI-driven decision support system that takes the guesswork out of crop health management.

Overview

A decision support system for everyday farming.

What it is

A platform that combines a Multi-Agent System, Agentic Retrieval-Augmented Generation, vector databases, and GIS-based visualisations to support farmers and agricultural officers throughout the crop lifecycle.

  • Multi-Agent System for coordinated reasoning across specialised concerns.
  • Agentic RAG grounded in verified agricultural data.
  • GIS-based visualisations for regional, macro-level insight.
What it solves

Crop losses caused by extreme weather and pests, inefficient use of water and inputs, and the lack of accessible agricultural expertise in rural communities.

  • Crop losses from extreme weather and pest outbreaks.
  • Uncoordinated use of water, fertilizer, and pesticides.
  • Limited access to agricultural expertise in rural areas.
System Architecture

A multi-agent system coordinated by an orchestrator.

Six specialised agents handle one concern each. An orchestrator routes requests, manages state, judges safety, and composes the final response.

Croptimum system architecture: users connect through a mobile and web frontend to a backend that hosts REST APIs, the community and heatmap services, and a RAG pipeline that reads from a vector database; an orchestrator with intent classifier, state manager, judge, and response synthesizer coordinates six specialised agents (Soil, Weather, Disease, Price, Pest, Planting) which talk to the Large Language Model.
Specialised Agents

Disease Agent

Identifies crop diseases from leaf imagery via Visual Question Answering grounded in PlantVillageVQA.

Pest Agent

Identifies insect pests and damage signs, grounded in the IP-VQA dataset.

Soil Agent

Interprets local soil context for fertility and irrigation guidance.

Weather Agent

Pulls live climate variables to ground recommendations in current conditions.

Price Agent

Surfaces commodity prices and trends so crop choice can factor in market value.

Planting Agent

Recommends crops and planting windows from water calendars and agro-met advisories.

Orchestrator

Four modules coordinate the agents end to end.

Intent Classifier
Decides which specialised agents a question needs.
State Manager
Tracks the conversation state across the multi-agent loop.
Judge
Checks every response for harmful or unverified output before delivery.
Response Synthesizer
Composes the final user-facing answer from agent outputs.
Request flow
01

Classify

Intent Classifier picks the specialised agents needed for the question.

02

Ground

Specialised agents gather evidence via Agentic RAG and live data sources.

03

Judge

Judge checks the draft response against safety and grounding rules.

04

Synthesize

Response Synthesizer composes the final answer for the user.

Technology Stack

The stack that powers Croptimum.

Each layer below names the technology Croptimum uses and what role it plays in the system.

App Flutter Flutter

Cross-platform mobile (and web) experience.

API FastAPI FastAPI

Asynchronous request handling.

Orchestration LangGraph LangGraph

Stateful multi-agent coordination.

LLM Gemini Gemini

Reasoning and response generation.

Vector DB Qdrant Qdrant

RAG retrieval over verified data.

App DB Supabase (Postgres) Supabase (Postgres)

Structured data and spatial mapping.

Object Storage AWS S3 AWS S3

Reference image storage.

Cloud AWS AWS

Hosting and scaling.

CI/CD GitHub GitHub

Build and deploy automation.

Observability Signoz Signoz

Live metrics and alerting.

Quality & Safety

How outputs are kept trustworthy.

Agentic RAG grounding

LLM responses are restricted to verified agricultural data stored in the vector database.

  • Retrievals come only from the verified vector database.
  • Image-similarity search adds visual context for diagnoses.

Judge Agent

A final-stage agent checks every response for harmful content before it reaches the user.

  • Reviews every completed response.
  • Blocks harmful or unverified output before delivery.

Evaluation metrics

Performance and accuracy are quantified against expert-verified answers.

  • ROUGE-L for semantic overlap, BLEU for fluency.
  • Token usage and response times tracked continuously.

Observability

API health and load behaviour are monitored continuously.

  • Signoz dashboards across all API endpoints.
  • Stress testing to verify resilience under load.
Accessibility & Access Control

Designed to be usable, and safe to use.

Accessibility

Built so users with varying levels of technical literacy can use the system without friction.

  • Voice input for hands-free use.
  • Minimalist UI for low-literacy users.
  • In-app guides and contextual tooltips.

Access control

Sensitive data and public surfaces are separated by role-based controls.

  • Role-based controls protect sensitive farmer profile data.
  • Public forums and expert portals stay isolated from private data.