The Phule series of digital farming products didn’t start as a series. It started, in 2004, with a mobile app that scheduled irrigation for a single research plot at MPKV Rahuri, based on soil-moisture readings we captured with a hand-crank moisture meter and typed into an Excel sheet every morning.
We hadn’t built a sensor yet. We didn’t own a cloud yet — nobody did. The “app” was a Windows Mobile utility that spat out a weekly irrigation schedule from a lookup table.
Twenty years later that same instinct — measure something, translate it into a decision a farmer can act on, get it into their hand — is still what we do. Everything else changed.
What actually matters when building for farms
Six things, in rough order of how expensive they are to get wrong:
1. Connectivity is a lie
Every deployment plan we ever wrote assumed 3G would be available at the field. It occasionally is. Most of the time, farmers are 30 minutes to two hours from the nearest usable tower, and the sensor is another hour from the farmer.
Design offline-first, sync opportunistically, and treat the last write as the source of truth. If you build for the demo (WiFi in a research pavilion), you will ship something that doesn’t work in the actual field.
2. The farmer’s second language is the app’s first language
Marathi first, then Hindi, then English. Not the other way around. The app title, the notifications, the button labels, the weekly SMS fallback — all of it.
Most of our engineers are Marathi speakers. That helps. But we still user-test with farmers who have never used a smartphone before, on the phones they actually own — not the flagships we buy for QA.
3. Solenoid valves rust, sensors drift, gateways die
Field hardware is not office hardware. We assume every deployed sensor will disagree with reality within twelve months. The system needs a calibration workflow that a field officer can run from an Android phone. The system needs a “this reading looks wrong” flag that a farmer can raise from the app. Both of those matter more than model accuracy.
4. Advisory is a scheduling problem, not a data problem
The Phule Irrigation Scheduler works because it tells you what to do this morning, before you make the wrong decision on your own. If it told you “your soil moisture curve is trending downward,” you’d close the app and irrigate anyway.
Reduce the model output to a specific action, with a specific quantity, at a specific time. Everything else is decoration.
5. Weather matters, but not the way you think
The nearest IMD station is usually 40+ km away, which for irrigation-relevant weather (evapotranspiration, wind, humidity) is basically noise. Our smart-weather-station rollout was expensive but paid for itself the first time the advisory recommended skipping an irrigation cycle because on-field data disagreed with the regional forecast.
6. Government-adjacent institutions move at their own tempo
MPKV is a university. Deployment happens across academic calendars, kharif and rabi seasons, agriculture-department budget cycles, and vice-chancellor terms. Plan releases around those, not around sprint reviews.
Things we’d do differently
Two things.
The first web dashboard we shipped tried to be a full LMS-style admin panel. Nobody used it. What people actually wanted was a WhatsApp bot that answered “should I irrigate today?” — and eventually that’s what we built alongside the app.
We over-invested in cross-platform mobile early. Ionic and later Cordova served us on projects where speed-to-first-farmer mattered more than long-term maintenance. But the pest-and-disease diagnostics app (PPDM) really wanted camera pipeline access we couldn’t easily get from hybrid frameworks, and we rewrote it in native Kotlin after two years. If we had started native there, we’d have saved money.
Where it goes next
We’re eight products deep into the Phule series and there’s a ninth that’s shipping this year. The domain hasn’t changed — farmers still need to know what to do this morning. What’s changed is the availability of on-device ML, cheaper LoRa sensor networks, and a class of large language models that are actually useful for agri-advisory Q&A in Marathi.
We’re not going to pretend that changes everything. Water still needs to be measured, soil still cracks in April, and the decision that matters is still the one that gets to the farmer before they irrigate.
But it does mean the ninth product is going to be built differently from the first eight. In a good way.
If you’re working on an agri-tech project and would like to talk, reach out.
