WORK · AI + EDGE · AGRICULTUREPILOT

AI + Edge Agriculture

Field sensors and edge inference feeding irrigation and crop-condition decisions, buffered through intermittent connectivity, with an agronomist approving every change before it reaches the field.

SYSTEM RECORDCS-02

INDUSTRY: Agriculture · irrigated field crops
SYSTEMS: Edge & Connected · AI Agents
SCOPE: One farm, two fields, one growing season
OWNER: Head agronomist (client)
STAGE: Assess → Pilot (current) → Engineer → Operate
MEASUREMENT: Agreed baseline; period in progress

photo: field sensors and edge unit on siteREAL OPERATION · MUTED · NO STOCK

Operational Context

A family-run farm growing irrigated field crops across two fields with different soil profiles. Irrigation is scheduled by the head agronomist from experience, weather forecasts and manual soil checks taken while walking the field.

The farm has patchy mobile coverage, no fixed network in the field, and a small team for whom any new system has to save time on day one or it will not be used.

Problem

Soil and canopy conditions were read by hand, at most once a day, in a few spots. Irrigation decisions lagged the weather by a day or more. When a decision was wrong there was no record of what had been seen or why it had been made, so nothing could be learned from it.

Business Consequence

Water and energy are spent on the schedule rather than the need. Under-watering in one zone and over-watering in another in the same week was common. The agronomist’s time was spent collecting readings rather than interpreting them.

COST: Water, pumping energy, agronomist hours
RISK: Yield variance between zones; stress events noticed late
EVIDENCE: None retained: decisions were not recorded

Constraints

The constraints shaped the architecture more than the technology did.

CONNECTIVITY: Intermittent 4G only. The system must keep working offline for days.
POWER: Solar and battery at the gateway; low-power sensors.
AUTHORITY: No automated actuation. The agronomist approves every change.
ADOPTION: One screen, in the field, usable with wet hands.
DATA: Farm data stays with the farm; aggregated only with consent.

Previous System

A paper notebook, a weather app, and a manual irrigation controller programmed weekly. The notebook held the knowledge; the controller held the schedule; nothing connected them.

06 · 07

System Design & Architecture

The system is one loop: sense the field, process locally, forward when the link allows, compare against the field’s own history, propose a schedule, and wait for the agronomist. AI is one bounded component that reads and proposes. It cannot actuate.

Edge-first by necessity: threshold alerts and data buffering run on the gateway so the farm has the same information on a day with no signal as on a day with full coverage.

ARCHITECTURE · CS-02

SENSORS
Soil moisture, canopy temp, weather
LORA · 15 MIN
EDGE DEVICE
Field gateway
EDGE LINUX · LOCAL RULES
TELEMETRY
Buffered uplink
MQTT · 4G · STORE & FORWARD
CLOUD / DATA
Field record · history
TIME SERIES · ASSET DB
AI AGENT
Irrigation proposal
READ · PROPOSE · NO ACTUATE
AGRONOMIST
Approves or edits
OWNER · ONE CLICK PAUSE
ACTION
Schedule to controller
LOGGED · OBSERVED

Human Responsibility

The head agronomist owns the system. Every irrigation change is approved by a person, every approval and edit is logged, and the agent can be paused from the same screen.

Who owns each decision, and what the system may do
DECISION	WHO	SYSTEM MAY
Change an irrigation schedule	HEAD AGRONOMIST	Propose a schedule with the readings and reasoning that led to it
Respond to a stress alert	AGRONOMIST ON DUTY	Raise the alert with location, trend and a suggested check
Pause or resume the agent	HEAD AGRONOMIST	Nothing. The pause is a human control.
Share aggregated data externally	FARM OWNER	Nothing without written consent

Engineering

Sensors report over LoRa to a solar-powered field gateway running Edge Linux. Local rules handle thresholds and buffering; MQTT with store-and-forward carries events to the cloud when the 4G link is up. Cloud services keep the field record and time series, and the agent runs against them on a schedule and on new data.

ASSESS: 3 weeks: field walk, interviews, sensor placement plan, baseline definition
PILOT BUILD: 6 weeks: sensors, gateway, telemetry, field record, agent, approval screen
PILOT RUN: One growing season, in progress

Reliability

Designed for the days the link is down.

OFFLINE: Gateway buffers up to 30 days of readings; local alerts continue
REPLAY: Ordered, deduplicated forwarding when the uplink returns
HEALTH: Battery, signal and heartbeat per device; alert to the agronomist if silent
FALLBACK: If the agent is unavailable, the previous approved schedule stands
SECURITY: Per-device credentials; TLS gateway to cloud

Measurement

A baseline was agreed before the pilot started. Results will be reported against it when the season closes; nothing is reported before then.

PRIMARY: Water applied per hectare vs. prior season, same crop
SECONDARY: Agronomist hours on data collection; stress events noticed within 24 h
QUALITY: Proportion of proposals approved without edit
BASELINE: Prior season records and the first four weeks of the pilot

Results

The pilot is in its measurement period. Figures will be published here against the agreed baseline when the season closes. Until then, only qualitative observations are reported.

PRIMARY OUTCOME
Measurement in progress
Reported at season close against the agreed baseline.
OBSERVED SO FAR
Decisions now have a record
Every proposal, approval and edit is logged with the readings behind it.
OBSERVED SO FAR
Field data survives the link
Multiple multi-day outages with no data loss.

Stacx24 does not publish percentages or ROI figures that have not been measured against an agreed baseline.

Current Status

PILOTSTATUS AS OF OCT 2026

Pilot live across two fields. The approval screen is in daily use. Hardening and the production handover are planned after the season closes and the results are reviewed with the farm.

PRODUCTION: Running in the client’s operation, measured, supported.
PILOT: One bounded system, live, with a baseline and an agreed measurement period.
EARLY DEPLOYMENT: In use, still being hardened; measurement period not yet complete.
TECHNICAL DEMONSTRATION: Built to prove an approach. Not operating a real workload.
CONFIDENTIAL CLIENT: Operating; client and figures withheld at their request.
FOUNDER-LED PRIOR WORK: Delivered by Stacx24 founders before the company; shown for method, not as company track record.

Technologies

Listed by layer. The system is the product; these are its components.

SENSE: Soil moisture probes
Canopy temperature
Weather station
EDGE: LoRa
Field gateway
Edge Linux
Local rules
TRANSPORT: MQTT
TLS
Store & forward
4G
CLOUD: Time-series store
Field record
Ingestion
INTELLIGENCE: Irrigation agent
Baseline comparison
Evaluation set
PEOPLE: Approval screen
Pause control
Audit log

NEXT STEP

Have a similar operation?

Describe it as it runs today. We will tell you whether this system shape fits, and what would be different for you.

Discuss a System All work