KisanAi

• 

  Home page

• 

  Image showing how easily crop suggestion can be drawn using queries

• 

  Image showing the multilingual capabilities of the agent for layman queries

• 

  Image showing the loans and subsidies available for farmers in their state

• 

  Image showing the voice input feature supporting differnt languages as input

• 

  Image showing the backend logs

• 

  Image showing features like irrigation assist forfarmers

• 

  image showing the fertlizers assist and selling advice and peak price for crops

🌱 About the Project: KisanAI 💡 Inspiration

Agriculture is often called the backbone of India’s economy, yet farmers face climate uncertainty, fragmented information, and low digital access. We were inspired by the question:

“What if every farmer could have a personal AI advisor that speaks their language, understands their needs, and helps them make smarter decisions?”

That vision led us to build KisanAI — a multilingual, voice-first Agentic AI assistant that empowers farmers with reliable guidance on soil, crops, weather, finance, and markets.

🛠️ How We Built It

Frontend: A responsive chat + voice UI built in HTML/JS with dark/light mode and audio playback.

Backend: Flask server (ai8.py) orchestrating ML models, LLM reasoning (Llama2 via Ollama), and multilingual translation.

Machine Learning:

Soil pH prediction models.

Crop classification and irrigation/fertilizer recommendation models.

Datasets: Cleaned and consolidated sources from ICRISAT, IMD, Agmarknet, and Government subsidy portals.

Knowledge Retrieval: Hybrid reasoning pipeline combining ML + rules + datasets, with LLMs used only for entity extraction to reduce hallucinations.

Deployment: Hosted on Google Cloud VM with demo and GitHub repo available.

📚 What We Learned

How to normalize messy public datasets (e.g., handling inconsistent units, aliases like “sarson = mustard”).

The importance of grounding LLMs with factual datasets for reliability in critical domains like agriculture.

Designing voice-first and multilingual systems that work for users with low literacy.

Balancing explainability with automation — farmers trust step-by-step instructions more than “black-box” answers.

🚧 Challenges We Faced

Data Quality: Many agricultural datasets had gaps, inconsistencies, or outdated values. We solved this with imputation, alias-mapping, and consolidation into master datasets.

Voice Recognition in Rural Contexts: Handling noisy environments and mixed-language queries (e.g., Hinglish) was tricky.

Infrastructure Limits: Cloud-hosted demo is slower due to limited GPU resources.

Scalability: Current version works for demos, but scaling to millions of farmers will require APIs, offline models, and IoT integration.

🚀 The Journey Ahead

We see KisanAI as more than a hackathon project — it can evolve into a national-scale advisory platform. Next steps include:

Offline mobile app with preloaded models.

Integration with Govt APIs (Agmarknet, IMD, PM-Kisan).

Image-based crop disease diagnosis.

IoT sensor integration for real-time soil/weather feeds.

A multi-agent ecosystem where specialized agents handle soil, markets, and finance, working together to serve farmers holistically.

✨ In short: KisanAI is where Agentic AI meets agriculture — empowering farmers, reducing risk, and shaping a more resilient future for food security.

Built With

• flask
• ollama
• open-meteo-forecast
• python
• scikit-learn
• xgboost