AI drift detection (on the roadmap)

Status: roadmap, in active development. What ships today is rules-based threshold alerting; drift detection adds a learned-normal layer on top. Design partners are welcome - the more sites we have learning data from, the faster the model is genuinely useful.

Threshold-based alerts are good at telling you that a fridge has breached its safe range. They are bad at telling you that a fridge is about to. Drift detection closes that gap by learning what each asset's normal behaviour looks like and surfacing assets that are deviating from themselves, before they trip a hard threshold.

What drift detection will do

The first release of drift detection is scoped to one job:

For every asset, surface the moments when its temperature behaviour is statistically unlike its own past behaviour - even if it is still inside the safe range.

That's deliberately narrow. We are not promising the model will diagnose the cause; we are promising it will tell you to look, with enough lead time to act.

Concretely, the dashboard will gain:

A "drift" badge on assets whose current behaviour is unlike their own baseline.
A "looking unwell" filter at the site overview level, sorted by how confidently the model thinks the asset is drifting.
A maintenance-class alert (lower urgency than an actionable threshold breach) for assets that have drifted past a configurable confidence band.

The intent is that maintenance gets booked from a drift alert, not from a stock-loss event.

Why this is hard to do well

Cold-chain assets are noisy. A fridge in a busy cafe at 8am behaves wildly differently from the same fridge at 11pm. Naive anomaly detection (e.g. "alert if outside +/- 2 standard deviations") fires constantly on door openings and is worse than no model.

To do this honestly we are working through:

Time-of-day baselines so the morning-rush warming is not flagged as drift.
Seasonality so the model doesn't think every Christmas Eve is anomalous because Christmas Eves are different from regular Mondays.
Per-asset learning rather than per-class baselines, because two display fridges of the same model can have very different normal behaviour depending on where they are in the room.
Confidence calibration so the alert says "70% confident this asset is drifting" instead of pretending to certainty.

Why we're confident we'll ship it

Every reading collected by every existing customer is feeding the training set. The platform has been logging telemetry continuously since launch, so by the time the model is ready to ship, the training data is real and grounded in real Australian operating conditions - not synthetic data, and not data from another country's climate.

The underlying time-series store is already designed for the kind of windowed queries the model wants; we are not boiling the architectural ocean to add drift detection - we are adding a feature on infrastructure already built to support it.

When it ships, you do not need new hardware

When drift detection ships, every existing customer with an active monitoring subscription receives it as part of the same dashboard and the same monthly HACCP report. No new SKU, no new hub, no upgrade. The hardware you buy today is the hardware drift detection runs on tomorrow.

This is explicit policy: AI features should not be a paywall on top of the data you are already paying us to collect.

What ships today (so you know the gap)

Today, the alert engine is rules-based:

Configurable per-asset temperature thresholds.
Configurable debounce windows so brief excursions do not page.
Configurable severe thresholds for escalation.
Mobile action pages and audit trails for every incident.

See alerts & escalation for what's live now.

Drift detection adds a maintenance-class signal on top of that engine; it does not replace it.

Become a design partner

If you operate at least three cold-chain assets and would like to influence what drift detection actually does:

Contact us and mention "drift design partner".
Design partners get earlier access to dashboard previews and direct conversations with the team building the model.
In return, your operational reality (which alerts you actually want, which you want to silence, which assets surprised you with how they fail) shapes the model.