Live Machine Learning Web Dev

KrishiBot — AI Agriculture Assistant

An offline-first AI agricultural assistant for smallholder farmers in South Asia — powered by a local LLM, bilingual (English & বাংলা), no cloud, no API keys.

Next.js FastAPI Ollama TypeScript Tailwind CSS SQLite Docker
KrishiBot home page with the heading Smart Agriculture Assistant Powered by Local AI

Overview

What it does

Detects crop diseases from photos, answers farming questions in natural language, plans treatments on a 14-day calendar, checks weather-driven disease risk, and lets farmers share disease sightings with the community.

Problem it addresses

Smallholder farmers often can't rely on cloud services. The LLM runs locally through Ollama — no subscriptions, no internet required after setup, and no data leaves the device.

Built as

A Final Year Project demonstrating how open-source LLMs can address real agricultural problems for users who can't rely on cloud services.

The backend is FastAPI with SQLAlchemy, talking to Ollama's qwen2.5:3b model for chat, analysis, advisory and treatment plans. Image analysis feeds filename-based crop hints to the text model, which keeps RAM requirements low; the qwen2.5vl:3b vision model is optional. The Next.js 14 frontend streams chat responses token by token over Server-Sent Events.


Key Features

Bilingual streaming chat

English and Bangla with an instant toggle, token-by-token SSE streaming, multi-session history and voice input via the Web Speech API.

Crop disease detection

Upload a leaf or crop photo for a structured diagnosis with severity staging (Early / Moderate / Severe), a confidence rating, and a PDF report export.

Market context

Estimates crop value, projected loss, treatment cost and net saving to give an ROI verdict on treating a disease.

14-day treatment calendar

Auto-generated day-by-day schedule per detected disease, with a robust JSON fallback if the LLM output is malformed.

Weather & risk alerts

Open-Meteo forecasts (no API key) drive heuristic fungal and bacterial disease-risk scoring over the next 48 hours.

Community disease board

Anonymous crowd-sourced disease reports aggregated across all 8 Bangladesh divisions, with a stats dashboard.


Tech Stack

KrishiBot technology stack
Frontend Next.js 14 (App Router), React 18, TypeScript, Tailwind CSS, Lucide icons
Backend FastAPI, SQLAlchemy 2, Pydantic v2, httpx (async)
LLM Ollama — qwen2.5:3b (text), qwen2.5vl:3b (vision, optional)
Database SQLite by default; PostgreSQL supported via environment variable
Services ReportLab (PDF reports), Open-Meteo (weather), Server-Sent Events (streaming)
Deployment Docker image for the backend; frontend on Vercel