Inspiration
India has 86 million small and marginal farmers. They produce the food. They bear the risk. And they consistently sell at the wrong mandi at the wrong time, not because they make bad decisions, but because they never had the information to make good ones.
My PhD research on post-harvest supply chain management in Karnataka revealed a specific, measurable failure. The gap between what a farmer receives at their nearest mandi and what they could have received at the optimal mandi, adjusted for transport cost, market depth, and timing, is routinely ₹150 to ₹300 per quintal. For a Farmer Producer Organization managing 500 farmers, that gap compounds into lakhs of rupees lost every harvest cycle. This is not a farming problem. It is an information architecture problem. And that is exactly what AI agents are built to solve.
What it does
Mandi Intelligence Agent is an agentic AI system that helps Farmer Producer Organizations decide where, when, and how to sell their crops. An FPO coordinator asks a natural language question such as:
“Should I sell 200 quintals of maize from Hassan today? Compare Hubli, Davangere, and Bangalore.”
The system returns a grounded, structured recommendation with net realization figures, transport-adjusted prices, arbitrage opportunities, and risk assessment.
The live demo query returns: Bangalore ₹2,119 per quintal vs Davangere ₹1,973 per quintal vs Hubli ₹1,675 per quintal, with the full reasoning for why Bangalore wins despite the longer distance. That is ₹88,800 more on a single 200-quintal transaction.
How we built it
The system uses two Amazon Nova models in a live end-to-end pipeline.
Amazon Nova Multimodal Embeddings indexes the knowledge base including price summaries, MSP policy documents, ethanol procurement rates, and crop advisories. At query time, the incoming question is embedded using Nova Embeddings at runtime, ranked against the indexed knowledge via cosine similarity, and the retrieved context is injected directly into the Nova Lite prompt.
Amazon Nova 2 Lite is the reasoning brain. It receives the retrieved semantic context alongside deterministic price and transport calculations, then generates the structured advisory: Recommendation, Reasoning, Key Numbers, Risks, and Timing. It does not invent prices or policy facts. It reasons over grounded data.
The architecture also includes Nova Act as the production data ingestion layer. It automates browser navigation of eNAM and Agmarknet portals that have no public API. The system also integrates Nova 2 Sonic as a Hindi voice interface for field coordinators. Both components are fully coded and architecturally integrated, ready for deployment in supported regions.
Challenges we ran into
The official Agmarknet API on data.gov.in was unreliable during development. We solved this by seeding a verified Karnataka price database directly from official datasets, keeping the agent grounded rather than dependent on a fragile external call.
The Nova model IDs changed between documentation versions, requiring careful verification of the correct payload schema for each model. Designing an agent that genuinely does not hallucinate prices, and refuses to invent numbers while only reasoning over what the database actually contains, required multiple prompt iterations to get right.
Accomplishments that we’re proud of
We built a working agent that does more than summarize data. It makes an actionable sell recommendation with transparent reasoning and quantified tradeoffs. The system compares multiple mandis, adjusts for transport, retrieves relevant policy and market context semantically, and returns a structured decision instead of a generic answer.
We are especially proud that the demo is grounded. The recommendation is not a hallucinated narrative. It is tied to deterministic calculations and retrieved evidence, which is the only reliable way to build high-stakes agricultural AI. The architecture is already designed for real-world expansion: dashboard today, automated ingestion with Nova Act next, and voice-first farmer access through Nova Sonic after that.
What we learned
The hardest part of building an agricultural intelligence system is not the AI. It is the grounding. An agent that confidently recommends the wrong mandi is worse than no agent at all.
The architecture of retrieval-augmented reasoning, where Nova Embeddings retrieves policy context and Nova Lite reasons over verified price data, turned out to be exactly the right pattern for high-stakes advisory systems.
We also learned that reliability matters more than theatrical complexity. In a real farm-gate decision workflow, trust comes from traceable numbers, not flashy language.
What’s next for Nova Market intelligence
Scale to all Karnataka commodities. Integrate Nova Act for nightly automated data collection. Deploy Nova Sonic for voice-first access in Hindi and Kannada for FPO coordinators who never touch a dashboard. Expand to Andhra Pradesh and Maharashtra mandi networks.
The information asymmetry that costs Indian farmers billions annually is solvable. This is the architecture that solves it.
Log in or sign up for Devpost to join the conversation.