● refactoringbuildingsamridhlimbu.com/projects/farmers-intuition · v0.1
Farmers Intuition
● hack48 · 48h 20hAI precision irrigation advisory platform built at Hack48 — billed as 48 hours, cut to 20 by the investor panel mid-event. Two-layer ML engine (prediction + recommendation) paired with a voice assistant named Sage — farmers ask questions in natural language and get actionable irrigation advice spoken back.
Context
Hack48, March 2026. The problem: farmers make irrigation decisions based on intuition and weather apps — neither accounts for actual soil moisture or crop state. Farmers Intuition replaces the guesswork with a two-layer pipeline: a prediction model that reads sensor and forecast data, and a language layer that turns the prediction into a spoken recommendation via the Sage voice assistant.
Timeline
Mar 2026 · h+0
Hack48 kickoff
Advertised as a 48-hour hackathon. Problem space: precision agriculture. Farmers make irrigation decisions based on gut feel and weather apps — neither is calibrated to soil or crop state.
Mar 2026 · h+2
Investors cut the clock
The investor panel reduced the build window from 48 hours to 20. Every scope decision after this point assumed less than a day of build time remaining.
Mar 2026 · h+6
Two-layer ML engine
Layer 1: prediction model (soil moisture, weather forecast, crop type → irrigation need). Layer 2: recommendation engine that converts the prediction into a plain-English action.
Mar 2026 · h+14
"Sage" voice assistant
ElevenLabs for voice synthesis. Farmers ask questions in natural language; Gemini handles intent and context; Sage speaks the answer back. Designed for use without looking at a screen.
Mar 2026 · h+20
Submission
FastAPI backend, two-layer ML engine, and voice assistant delivered inside the compressed 20-hour window.
Key technical decisions
01two-layer ml › single end-to-end model
Splitting prediction from recommendation keeps each layer independently improvable. The prediction layer can be swapped for a better soil model without touching the language layer.
02voice-first interface › dashboard ui
Farmers don't carry laptops in paddocks. A voice interface (ElevenLabs + Gemini) works on a phone in a pocket — ask a question, get an answer, keep working.
03fastapi › django / flask
FastAPI gives async-native endpoints and automatic OpenAPI docs. In a hackathon, the docs double as a team contract — frontend and ML layers agree on the schema from hour one.
04gemini › fine-tuned local model
20-hour constraint: no time to fine-tune. Gemini handles natural-language understanding out of the box; the ML engine handles domain prediction. Clear separation of responsibilities.
The two-layer pipeline
kairos/scheduler.pypy
1# layer 1: predict irrigation need
2prediction = model.predict({
3 "soil_moisture": sensor_reading,
4 "forecast_mm": weather_api_data,
5 "crop_type": profile.crop,
6})
7
8# layer 2: convert to natural-language advice
9advice = gemini.generate(
10 prompt=build_prompt(prediction, profile)
11)
Layer 1 handles domain knowledge — soil, weather, crop type. Layer 2 handles language. Neither layer tries to do both. In a 48-hour build this division mattered: team members could work on each layer independently against a shared schema.
Stack
BackendFastAPI · Python
AI / MLGoogle Gemini · custom prediction model
VoiceElevenLabs (Sage)