FarmHand

TL;DR: A $73 sensor kit that lets growers see every crop bed live and waters for them. Saves time, money, water and the crop.

Inspiration

Florida, spring 2026: the worst drought in 25 years.
A nursery in Bonita Springs watered morning, noon and night, and still threw plants away.
The real problem: nobody could see what the soil needed.
Growers pay 3 ways:
- hours walking beds with a probe
- water poured on soil that was already wet
- crops lost when a bed dries out unnoticed
Commercial sensors cost $1,200+ each, plus a yearly subscription.
We wanted the same visibility for the price of a few parts.


See it live	Both soil boxes in 3D with moisture and temperature, updated every second
Decides	Our decision model picks: water now, hold off, or wait for rain, and shows why
Acts safely	Rules on the chip: capped pours, spacing, a daily limit, never waters blind
Speaks	Tap Listen and an ElevenLabs voice reads the farm aloud
Plans ahead	Slide moisture and temperature to see which crops thrive, then send it to the model
Maps it	Miami-Dade farms, vegetation, live soil water and what could grow on any spot you tap

Box A = Farm Hand. Box B = a plain timer. Side by side, so the difference is obvious.

Hardware: ESP32 · 2 soil probes · 2 temp probes · 2 relays · 2 pumps · 12 V → LM2596
Firmware: C++, a loop that never blocks, safety rules on the chip
Server: a Mac mini runs the decision model and sends each call back to the board
Data: Tiger Cloud stores every reading as time series
Model: fine-tuned on an RTX 4070, tested on 21 months it never saw, decides in milliseconds
Frontend: React · three.js · Vite · Leaflet · a Blender 3D model driven by live data
Before parts arrived: full rig simulated in Wokwi, plus a virtual box in the browser

Google Gemini API: we built Farm Hand with Gemini in Google AI Studio. Gemini pulls, analyzes and plots our data: 1,347 farms across Miami-Dade (crop type right 87% of the time), the weather, and which crops fit each field.
Microsoft: AI in the experience, not a chat window. The model waters real soil on its own and growers use the crop simulator to do a real task.
ElevenLabs: one tap and a voice reads each box's moisture, temperature and what the model is doing. Hands-free in the field.
Tiger Data: every reading goes into a Tiger Cloud hypertable, with 1-minute and 1-hour rollups and compression, so charts stay fast.
Assurant: the model runs locally, so farm data never leaves our machine. Every call shows its reason and the water used. We chose a small model on purpose: milliseconds vs 64–180 s for a cloud agent.

Relay stuck ON: 3.3 V logic can't switch a 5 V relay off. Fix: release the pin for off, pull it to ground for on.
Boot pins blocked flashing (D2, D3, D12). Fix: moved the sensors.
Readings dropping out. Fix: a USB live view next to Wi-Fi found a loose power line.
Temp probe froze the loop for about 750 ms. Fix: non-blocking reads.
Bluetooth ate the Wi-Fi's memory. Fix: turned it off.
Cloud AI agent was too slow (64–180 s). Fix: a small local model.

56.2% less water than a timer (season replay, 21 months of real Miami weather)
0 hours of crop stress vs the timer's 12
94.1% of calls right on 7,524 decisions the model never saw
$72.94 in parts vs $1,200+ for one commercial sensor
Measured on the rig: 23.5 ml/s per pump (235 ml in exactly 10 s)
Safe by design: no pour unless the chip's rules agree

Seeing the soil is most of the value.
Small and fast beats big and slow in a control loop.
Safety belongs on the device.
Hardware humbles you: voltages, boot pins and loose wires beat us more than the AI did.