KisanSetu

Empowering Indian farmers via Confluent & Gemini—bridging FPOs with voice insights to maximize crop productivity and income.

Comment

Inspiration

Recently, I visited a relative who is a farmer, he sat me down for hours, asking me to navigate complex government portals to find critical market data that he couldn't access himself due to the literacy barrier. Watching him struggle to find information that wasn't personalized or real-time was where I realised that the data was there, but it was effectively invisible to him. That frustration was the spark for KisanSetu: I wanted to build a solution that replaces complex dashboards with a simple voice conversation, ensuring that millions of Indian farmers like him can finally access the real-time insights they deserve without needing a translator.

What it does

In India, Farmer Producer Organizations (FPOs) are the economic backbone for small farmers, but they suffer from a massive "latency gap." Market prices change by the hour, but information reaches farmers days later via hearsay. KisanSetu eliminates this delay, transforming the FPO from a static administrative body into a real-time digital nerve center, and the farmer from a passive recipient into an informed decision-maker.

The User Flow: A Conversation Powered by Streaming Data We built a bridge that connects the illiterate farmer to the cloud: The Ask: A farmer opens the app and asks naturally in their local dialect: "Is it a good time to sell my wheat?"

The Bridge: Gemini processes the voice input and uses the Confluent Model Context Protocol (MCP) server to query Confluent Cloud. It checks live Kafka topics for real-time prices and weather without hallucinating.

The Feedback: As the farmer gets a voice answer, their query is produced back into a Kafka topic, instantly updating the FPO’s dashboard with live demand data.

KisanSetu provides the farmer with-

-Zero Literacy Barrier: Voice-First interface allows farmers to access enterprise-grade intelligence without reading complex charts. -Income Protection: Real-time price comparisons prevent "distress selling," ensuring farmers don't lose 30% of their income due to stale data. -Instant Adaptability: Farmers receive weather warnings from the FPO within seconds (e.g., 2:01 PM) of them being issued.

For the FPO - -The "Live Pulse": Admins see a real-time heat map of farmer concerns. If 50 farmers ask about "pest control" in one hour, the FPO spots an outbreak instantly. -Data-Driven Logistics: By monitoring query volume (e.g., a surge in "Soybean price" checks), FPOs can predict harvest supply and prepare logistics in advance.

How we built it

We architected KisanSetu as a modern, event-driven application that merges the immediacy of Voice AI with the reliability of enterprise data streaming. Our system creates a continuous feedback loop between the farmer's voice and the FPO’s data.

  1. Confluent Cloud & Kafka We established a robust event streaming backbone using Confluent Cloud (Apache Kafka). To move beyond static databases, we created dedicated Kafka topics for distinct data streams: market_activities for real-time interaction logs, weather_updates for hyper-local forecasts, and gov_mandi_prices for live market rates. To bridge the gap between legacy systems and our modern app, we implemented Confluent Source Connectors. This allowed us to ingest data from existing FPO SQL databases directly into our Kafka topics, ensuring the AI always had access to the ground truth. We also utilized the Confluent Schema Registry to ensure data consistency across all producers and consumers.

  2. The Bridge: Confluent MCP Server The biggest challenge in Generative AI is hallucination. To solve this, we integrated the Confluent MCP Server using the Model Context Protocol (MCP). This acts as a universal translator, giving our Gemini Agent the ability to use tools. When a farmer asks a question, the MCP server interprets the intent and routes the query to the correct Kafka topic or Flink SQL statement. This enabled standardized message consumption and production, allowing the AI to read market trends and write farmer queries back to the stream instantly without guessing.

  3. The Brain: Gemini Multimodal Live API For the farmer experience, accessibility was the priority. We integrated Google’s Gemini Multimodal Live API to build a voice-first interface. We bypassed traditional text inputs entirely; the app handles real-time speech-to-text and text-to-speech natively. The AI constructs context-aware responses by dynamically pulling data from the MCP bridge. If a farmer asks about wheat prices, Gemini retrieves the latest message from the gov_mandi_prices topic before speaking the answer.

  4. The Command Center: React & Real-Time Visualization For the FPO administrators, we built a high-performance frontend using React, TypeScript, and Vite. We utilized Recharts to visualize the streaming data. Because our architecture is event-driven, the dashboard updates live as messages flow through Kafka. Tailwind CSS and Lucide React were used to create a clean, accessible UI designed to perform efficiently on both mobile and desktop devices.

Architecture Overview: FPO Database → Confluent Source Connector → Kafka Topics → Confluent MCP Server → Gemini AI Agent → Farmer (Voice)

Challenges we ran into

Challenges we ran into

  1. Real-Time Data Synchronization: Challenge: Ensuring data consistency across multiple Kafka topics while maintaining low latency Solution: Implemented proper partitioning strategies and used Confluent Schema Registry for data validation

  2. MCP Server Integration: Challenge: Bridging the gap between Kafka streams and AI agents through the Model Context Protocol Solution: Built a custom MCP server that translates Kafka messages into AI-consumable context

  3. Voice Processing Latency: Challenge: Minimizing delay between farmer's voice input and AI response with real-time data Solution: Implemented streaming responses and optimized Kafka consumer configurations for minimal lag

Accomplishments that we're proud of

  1. Seamless Real-Time Integration: Successfully created a system where FPO database updates appear in farmer dashboards within seconds, demonstrating true real-time agricultural intelligence.

  2. Voice-First Agricultural AI: Built the first voice-activated AI assistant specifically designed for farmers, capable of understanding agricultural terminology and providing contextual responses using live market data.

  3. Scalable Architecture: Designed a system that can handle thousands of FPOs and millions of farmers through Confluent Cloud's distributed streaming platform.

  4. Production-Ready Integration: Implemented actual Confluent Cloud connectors and MCP servers that can be deployed in real agricultural environments, the app is also deployed using Cloud Run.

What we learned

This was our first time experiencing confluent and MCP servers both so learning them from scratch was a very fulfilling journey. We used a lot of tools like the Gemini CLI , played around creating different variations of the prototype using Google Antigravity and AI studio. Learning about new technologies like confluent - connectors , clusters will allow us to move forward with a lot of valuable and practical knowledge ahead which we are very grateful for.

What's next for KisanSetu

We want to evolve KisanSetu from a market intelligence tool into a comprehensive agricultural operating system. Our next steps focus on expanding accessibility through offline-first mobile apps and broader regional language support (Hindi, Punjabi, Marathi), ensuring every farmer can connect regardless of connectivity. Technically, we aim to deepen our "Neural Bridge" by ingesting IoT data from soil sensors and drones directly into Confluent, while integrating financial services and blockchain-based traceability. Ultimately, we envision scaling across India and the Global South, enabling farmers to access carbon credit markets and full-cycle AI crop management.

Built With

Share this project:

Updates