KrushiMitra explores how a farmer-facing yield tool could collect district, crop, rainfall, and temperature context without demanding specialist hardware or a desktop workflow.
The strongest part of the current project is the product and service architecture: a mobile React interface, typed components, and a FastAPI backend prepared for a future validated model. The repository currently uses mock prediction behavior, so this case study intentionally makes no accuracy or agronomic-impact claim.
Tech Stack
How the System is Built
Curated historical records → leakage-safe temporal split → baseline comparison → calibrated uncertainty → agronomic review.
Typed form state → API request → response card with a predicted range, historical comparison, and explanatory factors.
How It Works — Step by Step
Farmer Fills Input Form
The farmer selects their district, soil type (Clay Loam, Sandy, Black, Red), crop (Rice, Wheat, Maize, etc.), expected rainfall range, and growing season (Kharif / Rabi / Zaid). Designed for low-literacy environments with visual icons.
District · crop · rainfall · temperature · seasonal context
Validate a typed request
FastAPI receives the scenario, validates numeric ranges and categorical inputs, and converts the frontend payload into a stable service contract.
React form → Pydantic schema → scenario service
Return an illustrative scenario
The current public implementation uses mock prediction logic to exercise the end-to-end interaction. It is an interface prototype, not a validated agricultural model.
Current status: product prototype · benchmark pending
Contextualised Output
The response design groups the estimated range, historical comparison, and key factors so a future validated model can explain its output without changing the user flow.
{ estimate, range, comparison, factors, model_status }
Challenges & Solutions
Prediction claims arrived before validation
The interface made it easy to present model-like output before a trustworthy dataset, temporal evaluation, and calibrated uncertainty were in place.
Made model status explicit
The portfolio and case study now label the project as a prototype. A future promotion requires a documented dataset, reproducible baselines, and out-of-time validation.
Complex inputs for a broad audience
Weather and crop context can quickly become an intimidating form, especially on a small screen.
Mobile-first component flow
Used React, TypeScript, Radix primitives, and constrained controls to keep the scenario workflow responsive and understandable.
A future model must generalize across time
Random splitting would overstate performance when weather patterns and agricultural practices drift across seasons.
Define a temporal benchmark first
The next research milestone is an out-of-time comparison against district and crop baselines, followed by uncertainty calibration and domain review.