Agri-Agent Nexus: Building a Multilingual AI Advisor for Farmers in Their Own Language

Agri-Agent Nexus: Building a Multilingual AI Advisor for Farmers in Belagavi

The problem

Farmers in and around Belagavi, Karnataka, rely heavily on informal advice, guesswork, and delayed information when making decisions about crops, weather, pests, and market prices. Most existing digital advisory tools are built in English, which excludes a large share of the farming population who are more comfortable in Hindi, Kannada, or Marathi. That language gap not a lack of information, but a lack of accessible information is the core problem Agri-Agent Nexus set out to solve.

The solution

Agri-Agent Nexus is a production-grade agricultural advisory system that lets farmers ask questions and get answers in their own language — English, Hindi, Kannada, or Marathi. Instead of a static FAQ or a single-shot chatbot, it's built as a proper retrieval-augmented system so that answers are grounded in real agricultural knowledge rather than the model's guesses.

Under the hood:

• Gemini 2.0 Flash powers the core reasoning and multilingual response generation.
• Vertex AI Vector Search handles retrieval, so responses are grounded in relevant agricultural knowledge rather than hallucinated.
• Cloud Run serves the application, keeping it scalable and stateless.
• Firestore stores structured data and conversation context.
• Pub/Sub handles asynchronous event processing.
• Memorystore provides low-latency caching for faster responses.
• Secret Manager keeps credentials and API keys secure.

Why this isn't just a demo

A lot of hackathon-style AI tools stop at "it works when I demo it." Agri-Agent Nexus was built with the discipline of a real product: it has a 102+ test suite covering the system's behavior, and a RAGAS evaluation suite to measure retrieval and generation quality — not just whether the app runs, but whether the answers it gives are actually good.

What I learned

Building this reinforced a few things:

• Multilingual support is not just translation it's making sure retrieval and grounding work equally well across languages, not just the output text.
• Production-readiness (testing, evaluation, secrets management) is what separates a prototype from something a farmer could actually rely on.
• Local problems need local-language solutions — accessibility is often the biggest lever for real-world impact, more than raw model capability.

Try it / learn more
Github Link: https://github.com/Aditi25369
Linkedin : https://www.linkedin.com/in/aaditi-r-burse-744bb52a3/

Built as part of my ongoing work in GenAI/ML engineering, using Google Cloud's AI stack to make agricultural advice accessible to farmers in their own language.