Krishi-Sahayak

Farmers face erratic weather & complex schemes. KrishiSahayak uses AI to give crop advice, find loans & subsidies, and auto-fill forms—making farming smarter, safer & more profitable.

Comment

Inspiration# KrishiSahayak: AI for Smarter Farming and Finance

Inspiration 🌱

“Agronomy advice is tough; money is tougher.”
In rural India, farmers battle erratic weather, declining soil fertility, and confusing government portals. Subsidies, insurance, and low-interest loans exist, but most smallholders miss out because the system is too complex.

That problem inspired us to build KrishiSahayak: an agentic AI that provides crop advice + financial access in one workflow.

What We Learned 🎓

  1. Grounding is vital — farmers trust answers with citations.
  2. Multilingual ≠ translation — dialect handling is key.
  3. Agents need structure — function-calling kept outputs consistent.
  4. Design for reality — offline support and WhatsApp reminders matter.
  5. Measure actions, not words — we tracked form downloads and reminders, not just accuracy.

How We Built It 🔧

  • Granite LLM: multilingual dialogue, tool use, document drafting.
  • ADK: defines multi-agent workflow.
  • FinanceScout: web search for KCC, PM-KISAN, PMFBY, NABARD schemes.
  • DocGen: generates pre-filled PDFs & checklists.
  • ComplianceAgent: enforces guardrails, redacts PII.

Architecture

Farmer Input → IntakeAgent → AgroAdvisor → FinanceScout → DocGen → ComplianceAgent

What it does

KrishiSahayak is an AI-powered companion for farmers in India.
It analyzes soil, crop, and weather data to suggest actionable farming plans.
Our FinanceScout agent searches official portals for loans, subsidies, and insurance.
It auto-fills application forms and sends reminders via WhatsApp or SMS.
The result: smarter farming combined with easier access to financial support.

How we built it

We used IBM Granite LLMs for multilingual advice and reasoning.
The ADK framework let us orchestrate multiple agents into one workflow.
OCR and ASR pipelines helped process soil reports, images, and voice input.
FinanceScout integrated web search tools to fetch live scheme information.
Finally, DocGen created pre-filled forms and checklists for easy submission.

Challenges we ran into

Government portals had inconsistent formats with PDFs, tables, and scanned docs.
We faced noisy, code-mixed speech in farmer inputs during ASR processing.
Financial advice needed strict guardrails to stay safe and grounded.
Eligibility rules and deadlines varied widely across states and banks.
Balancing live web search with low latency was a constant struggle.

Accomplishments that we're proud of

Accomplishments that we're proud of

We designed a clear end-to-end architecture for farming and finance support.
Our workflow unites crop planning with financial scheme discovery.
We created pre-filled form templates tailored for KCC and PMFBY.
The system design supports multiple Indian languages and dialects.
Most importantly, we shaped a solution farmers can realistically use.

What we learned

Grounded answers with sources build trust far more than fluent text.
Multilingual AI must handle dialects, synonyms, and code-mixed speech.
Agent orchestration works best when each agent has a narrow, clear role.
Evaluation should track real actions like form downloads, not just accuracy.
Designing for offline and mobile-first use is critical in rural contexts.

What's next for Krishi-Sahayak

Build a functional prototype farmers can test on the ground.
Expand language support to cover more Indian states and dialects.
Add deeper integration with state-level portals and co-op databases.
Develop a lightweight Android app with offline features and voice I/O.
Pilot with local cooperatives to validate usability and real-world impact.

Built With

  • adk
  • and-the-ibm-adk-framework-to-connect-multiple-agents-together.-for-text-recognition-we-used-ocr-(tesseract)-and-for-voice-we-used-speech-to-text-(wav2vec2)-so-farmers-can-talk-or-upload-scanned-documents.-data-is-stored-in-postgresql
  • api
  • apis
  • fastapi
  • generation)
  • granite
  • imd
  • llms
  • ocr
  • openweather
  • postgresql
  • python
  • redis
  • reportlab
  • speech-to-text)
  • tesseract
  • wav2vec2
  • with-fastapi-to-serve-the-backend.-we-used-ibm-granite-llms-for-multilingual-advice-and-reasoning
Share this project:

Updates