Research and Development

User Research

Our development process began with direct conversations with farmers across multiple regions in Tamil Nadu, Kerala and Maharashtra to understand how agricultural decisions are made in practice. These discussions explored crop selection, weather-related risks, market considerations, information sources, and the challenges farmers face when interpreting available data.

Key findings from initial interviews indicated that weather uncertainty, market volatility, and access to actionable advisories are among the most significant challenges affecting agricultural decision-making. These insights continue to guide the design and priorities of the AgriLens platform.

Building an integrated, farmer-centred agricultural decision-support system

AgriLens brings together agricultural data, machine learning, language models and conversational interfaces into a single decision-support platform. Our research focuses not only on developing these individual components, but on understanding how they can work together to deliver information that is accurate, timely, understandable and useful in real agricultural contexts.

Data & Information Infrastructure

Agricultural information relevant to farmers is distributed across multiple sources, including meteorological services, agricultural advisories, market information, soil and land resources, and government and KVK publications. AgriLens is designed to bring these heterogeneous sources into a common data pipeline.

Our ingestion layer combines APIs, scheduled data retrieval and document processing pipelines. Structured data can be cleaned, validated and transformed before being stored for downstream modelling, while agricultural bulletins and other documents are processed into structured information for the advisory system.

Meteorological data
IMD weather and rainfall data, Agromet information and other weather sources.

Market data
Agricultural market prices and price trends.

Soil & land data
Soil characteristics and relevant land information.

Agricultural knowledge
Agromet bulletins, KVK advisories and government agricultural publications.

Machine Learning and Decision Support

The machine-learning component explores whether historical agricultural, environmental and market information can be used to provide useful crop-level decision support. The system combines relevant features such as weather conditions, market information, soil characteristics and crop context to generate comparative crop-level outputs.

Our approach is deliberately designed as a decision-support system rather than an autonomous decision-maker. Model outputs are intended to provide additional information to the farmer, while keeping the final agricultural decision with the farmer

We are experimenting with classical machine-learning approaches, including tree-based models, and evaluating them using held-out data and cross-validation. Model development is treated as an iterative process: new field and historical data can be incorporated as the system matures.

Explainability

Where appropriate, model explanations are used to identify which input factors contributed to a prediction. This is important for a farmer-facing system because a recommendation should ideally be accompanied by an understandable indication of why it was generated.

VAANI : Vernacular AI for Agrometeorology and Nowcasting in India

VAANI is the conversational layer of AgriLens, designed to make agricultural information accessible through natural voice interaction. Rather than requiring farmers to navigate multiple menus or search through documents, the system allows them to ask questions conversationally and receive information relevant to their context.

VAANI is designed to combine the farmer's profile and context with relevant weather information, agricultural advisories and AgriLens decision-support outputs. This allows the same conversational interface to support both specific advisory queries and broader agricultural questions.

The voice interface is being developed as an accessibility layer over the underlying information system rather than as an independent source of agricultural knowledge

Systems and Software Development

AgriLens is being developed as a modular software system so that individual components—including data ingestion, machine-learning models, advisory processing and conversational services—can be developed and evaluated independently while remaining integrable through APIs.

Backend

A FastAPI-based backend provides programmatic access to prediction, weather, market, advisory, user and conversational services.

Frontend

The farmer-facing application is made o flutter, designed around a simple workflow: onboarding → farmer profile → dashboard → crop information → recommendations → advisories → conversational assistance.

Data & storage

Structured data, raw documents, processed advisories and application data are stored separately according to their function, allowing the system to retain source information while maintaining structured representations for downstream applications.

Cloud infrastructure

Cloud infrastructure provides the storage, deployment, API and service layer required to run the system beyond a local development environment. The architecture is designed to support scheduled data ingestion, model inference, advisory processing and application requests as independent services.

Research and Evaluation Methodology

Our evaluation framework is designed to distinguish technical correctness from practical usefulness. A system can produce a technically valid response without necessarily producing information that a farmer can understand or act upon.

We are developing component-specific evaluation methodologies because different parts of the system have fundamentally different objectives. Rather than relying on a single overall accuracy measure, we evaluate the prediction, advisory, conversational and user-facing components independently.

We aim to evaluate AgriLens not only by model performance, but by whether the information produced is accurate, understandable and useful in real agricultural contexts.