Jandec Farms: Grow at home. Learn with AI. Automate with purpose. Every drop justified.

Jandec Farms is family-owned, family-operated, and very enthusiastically family-fed. Ginger, peppers, tomatoes, and cut flowers grow across five beds of roughly thirty square feet each sitting on a modest 1.25 acre plot. Every harvest ends up on the kitchen counter, dinner plate, or dining-room table. Nothing is sold, shipped, or measured against a revenue target.

The entire organisation chart is also remarkably efficient: one person.

Farm owner, grow-table operator, process-improvement lead, data analyst, IoT engineer, and AI adoption manager all report to the same individual — which makes performance reviews awkward, but accountability unusually clear.That became part of the experiment.The farm provided a deliberately small environment in which to apply ideas developed through the Mini-MBA in Artificial Intelligence: observe a real process, identify where decisions rely on habit rather than evidence, collect better data, remove unnecessary work, introduce automation where it adds value, and decide very carefully where automation should stop.

That changed the design brief.

A commercial grower can justify irrigation automation through labour savings, water reduction, yield improvement, or lower operating cost. At this scale, those arguments become almost comical. The financial return on optimising thirty square feet of ginger is unlikely to impress an investment committee.

Removing that justification, however, exposes a much more interesting one:

That is where the project stopped being simply about irrigation.

Soil moisture at root depth is invisible. A bed that looks dry may still contain plenty of water below the surface. A bed watered every Tuesday at 7:00 AM will happily receive water whether it needs it or not. Both approaches can be wrong, and both can remain wrong quietly for days.

So instead of beginning with automation, Jandec Farms begins with evidence.

Sensors create observations. Data turns those observations into trends. Analytics provides context rather than relying on a single reading. The decision layer evaluates whether watering is actually justified. Every recommendation retains the measurements and reasoning that produced it. And before water flows, a human remains accountable for authorising the action.

The objective is therefore not maximum autonomy. it is useful autonomy with boundaries.

Optimise the process where repetition adds no value. Use analytics where human observation is weak. Let AI connect signals that would otherwise be tedious to interpret. But keep the system explainable, traceable, and deliberately constrained wherever a wrong decision has consequences.

Because the interesting question is not whether AI can turn on a pump.That part is easy, The interesting question is whether it can earn the right to.

The five beds [As of 8/14/2026]

The objective was never "automate the watering". It was to build a closed decision loop in which the garden is continuously measured, those measurements are converted into evidence, the evidence is evaluated against known conditions, and the resulting recommendation arrives with something most automated systems conveniently leave out:

how sure are you?

Every cycle produces a measurable, recordable, explainable, and auditable decision. The system can recommend watering, recommend waiting, flag uncertainty, or decide that the available evidence is simply not good enough to make a responsible call.

What it cannot do is quietly turn on the water because an algorithm felt confident. That boundary is intentional. The machine owns the analysis. The human owns the consequence.

What the system decides

Every cycle, each bed receives exactly one of five verdicts. There is no sixth, and no free-text alternative:

Why this is agentic, not merely automated

A timer is automation. At 7:00 AM, the timer wakes up, opens a valve, and considers its intellectual contribution to agriculture complete.

Jandec Farms operates differently. The system repeatedly executes its own perception → reasoning → recommendation → human action → observation loop against one bounded goal: determine whether the available evidence justifies watering a particular bed.

It consumes current measurements, compares them with recent behaviour and historical baselines, evaluates drying patterns, determines which defined rule the evidence supports, selects one of five bounded verdicts, and attaches a confidence level to that judgement.

Then it stops. That boundary is deliberate. The system has authority to assess, recommend, and learn — but not to act. It cannot redefine the objective, expand its scope, or convert confidence into execution. Its autonomy is intentionally bounded to judgement; operational authority remains with the human.

The system owns the recommendation. The human owns the decision.

The non-negotiables

• No schedule gets a vote. Time can inform the decision; measurements have to justify it.
• Show the receipts.Every recommendation must expose the data, rule, trend, and confidence behind it.
• Human authority is absolute.The system may reason with conviction; the valve still answers to a person.
• Overrides are evidence, not embarrassment. Disagreement is recorded, observed, and learned from — not quietly deleted.
• History stays honest. Readings, decisions, actions, and outcomes are append-only; even the tomatoes get an audit trail.

The system is local-first. Sensing, reasoning, approval and actuation all happen on hardware inside the house; the internet is an outbound mirror that can be read but cannot reach in. Nothing in the cloud can water anything.

The hops, and what runs at each

Architecture — the full end-to-end poster click to expand

The whole system in one view: the lanes from soil to operator, every stage between them, and the feedback edge that turns a one-shot recommendation into a loop that improves. Open it full screen to read the stage labels.

Click to zoom & drag

Two invariants hold the shape of this picture in place, and neither may be optimised away: the system keeps working with the internet unplugged, and no cloud component can cause water to flow. Everything in section 6 is measured against those two lines.

Confidence is calculated fresh for each decision across five governance checks, with every answer recorded alongside the verdict. No gut feel, no “AI says 92%” magic. Five questions go in. One confidence number comes out. And when the evidence gets weaker, the confidence is expected to drop with it.

Everything below is read from the live archive, not from a projection. Where a goal is not yet evidenced, it is marked as not yet evidenced.

How the 620 decisions fell

Two-thirds of all cycles ended in the system either declining to advise or advising nothing. For a first season on partial hardware, that distribution is the right shape — an adviser that recommended watering two-thirds of the time would be telling us about its thresholds, not about the soil.

Success metrics, honestly scored

Why plain mathematics, and no machine-learning model

There is no machine-learning model in the current system, and that is intentional.

With one season of data and a handful of beds, an ML model would be learning more from noise than from truth. It might look sophisticated, but sophistication is not the goal. A decision that can be explained and challenged is far more useful here than one that is merely impressive.

So the system uses simple, established mathematics instead: temperature and humidity to derive vapour pressure, least-squares slope to understand drying rate, standard deviation to measure how stable a baseline really is.

The current win: The system is not trying to be clever; it is trying to be trustworthy.

How far this can grow

The system was built for a family garden, and at that scale it is doing what it was meant to do.

• Beds. Five beds are registered today, three are actively sensed, and extending coverage is mostly a matter of adding sensors and configuration rather than changing the architecture.
• Nodes. Sensor identity is per-channel, not per-board, so a second node in a far corner of the garden joins by adding rows to configuration files. Nothing in the ingest path assumes one device.
• Storage.Remains comfortably manageable with append-only files.
• People. Family access is intentionally narrow, secure, and sufficient for checking the garden remotely.

The more interesting scalability question is not how many beds it can monitor. It is how much responsibility should the agent eventually be allowed to carry. Today, Agentic AI owns the observation, reasoning, confidence, recommendation, and learning loop. The human still owns the final action. That boundary is the most important success of the current system.