AI agent

How the AI gets better every week

PlantHub tracks every action's outcome — before vs. after sensor readings, your feedback, weather context — and grows its confidence through four explicit tiers. Here are the actual thresholds, not marketing handwaving.

The four levels

NO_DATA

< 5 samples

You just claimed the device. The AI uses the seeded plant profile and conservative defaults. It explains every action it takes because it has nothing learned yet.

LEARNING

≥ 5 samples · ≥ 40% confidence

Enough actions have been observed that the AI starts to detect patterns — typical watering durations, response time of moisture sensors, time-of-day effectiveness. Confidence is rising but not high enough to fully trust.

ADVANCED

≥ 20 samples · ≥ 70% confidence · ≥ 60% consistency

The AI has multiple weeks of data, has a tight estimate of optimal parameters, and is producing consistent good outcomes. Recommendations carry stronger weight, and the AI starts proposing rule changes proactively.

EXPERT

≥ 50 samples · ≥ 85% confidence · ≥ 70% consistency · ≥ 70% effectiveness

The AI knows this specific plant in this specific environment. It can shorten or skip the LLM call for routine decisions, lean on learned thresholds, and surface cross-actuator patterns (e.g. "shade + mist drops temp 40% faster than either alone").

What "confidence" actually means

Confidence is a single 0-100 score combining four weighted factors:

Effectiveness (35%) — how often actions produced the expected sensor change (HIGH / MEDIUM / LOW / NEGATIVE outcomes).
Sample size (25%) — number of recorded outcomes, with logarithmic diminishing returns.
Consistency (25%) — how stable change percentages are across runs (coefficient of variation).
Recency (15%) — exponential decay favoring recent data.

Confidence is computed per actuator per device, then aggregated to the zone level for cross-device pattern detection.

Why levels matter

At NO_DATA and LEARNING, the system biases conservative — shorter waterings, longer cooldowns.
At ADVANCED and EXPERT, the system trusts learned thresholds and can act with reduced LLM involvement, saving cost.
If you replace a sensor or change plant species, the relevant confidence resets — you do not inherit stale learning from a different plant.

Where to see learning state

Every device's detail page surfaces the current confidence level, recent outcomes, and the AI's pending recommendations. The "AI insights" panel shows learned parameters with timestamps so you can see exactly what the system thinks the right setting is, and why.