From soil probe to shipped model: the whole stack, honestly labelled.
AgroSense is a solar field pod, a long-range radio link, an on-hub decision layer, and an automated pipeline that retrains that layer every night. Here is what runs today, what is bench-validated, and what is scoped for production.
AgroSense: one enclosure, three sensing domains Live · prototype hardware
A bench-validated prototype, live today at the live pod dashboard. It reads soil, air and a visual frame on every cycle, and closes the loop with its own pump.
8-in-1 soil probe
Modbus RTU probe reading moisture, temperature, EC, pH, nitrogen, phosphorus, potassium and salinity from a single insertion point.
Air quality: Sensirion SEN54
Temperature, relative humidity, PM1/PM2.5/PM4/PM10 and VOC index, sampled alongside the soil reading each cycle.
ESP32-CAM imaging
An on-board camera captures a field frame per cycle: the raw input SproutID's plant-ID pipeline is trained against.
Closed-loop pump control
The pod drives its own pump and meters litres per cycle, so every watering event has a measured volume attached, not an assumption.
Solar power
No mains connection required: the pod runs from its own solar supply, matching the off-grid reality of most field sites in the region.
LoRa radio
An EBYTE E22-900T30S module gives the pod km-class range back to the hub by module spec, field range to be validated, without WiFi or cellular coverage.
LoRa, because field sites don't have WiFi Live
Each pod talks to a LilyGo hub over an EBYTE E22-900T30S LoRa module, a link designed for km-class range at low power by module spec, exactly the profile a scattered farm or estate needs. No mains wiring and no local network are required between pod and hub.
Six decision-forest models, running on the hub itself Live
Decisions are made where the pod is, not in a distant data centre. The hub holds six trained models and answers each reporting cycle without waiting on a round trip.
Irrigate classifier
Decides whether this cycle's soil and climate readings call for water at all.
Duration regressor
If irrigation is called for, estimates how long the pump should run this cycle.
Fertigation classifier
Decides whether this cycle's NPK readings call for a nutrient dose.
Nitrogen dose regressor
Estimates the nitrogen dose for this cycle when fertigation is triggered.
Phosphorus dose regressor
Estimates the phosphorus dose for this cycle when fertigation is triggered.
Potassium dose regressor
Estimates the potassium dose for this cycle when fertigation is triggered.
Retrained every night, with no one at a keyboard Live
The models above don't stay static. An automated job keeps them current with what the field is actually reporting.
Every night at 06:00 UAE time, the pipeline pulls the latest real field readings from the cloud database, retrains all six on-hub models, and publishes the updated binaries. The hub checks for new binaries on its own schedule, roughly every six hours, so the loop closes without a person in it.
What ships to a pilot site Scoped & costed
The bench pod proves the concept. Production hardware is scoped and costed for the pilot batch: a real bill of materials, not an aspiration.
See what the hardware feeds.
The edge models above feed the programme-year AI layer and the live dashboard.