Most people have never had cause to think about how you would digitise a smell. I hadn't either.
A camera turns light into numbers and there are decades of tooling for what to do next. Smell has no equivalent. There is no universal unit and no obvious axis, and the same compound at two concentrations can read as pleasant or as a warning. Altered Carbon builds graphene sensors that respond to gases, but a sensor response on its own tells you almost nothing — it is a wobbly line. The value is entirely in knowing what that particular wobble means.
That is what ScentStudio is for.
What the platform does
Devices are registered, provisioned, monitored and updated in the field, with firmware and OTA handling and health telemetry, so a degrading unit is visible before anyone notices its readings have gone strange.
Training nodes are the interesting part. To recognise a smell you first need many labelled examples of it under known conditions. Those nodes gather that data and the platform turns it into datasets that can actually be trained against — which is considerably more of the work than the modelling.
Experiments run on top: batches, conditions, results, comparisons. Anomaly detection watches for readings that don't belong. All of it is scoped by organisation with role-based access, because one customer's scent data is commercially sensitive and emphatically not another's.
Dashboards turn it back into something a human can read — gauges and live state pushed over websockets, because someone tracking drift across a week should not have to keep reloading a page.
Building it twice
The first generation taught me what the product actually needed, which was not
what its architecture had assumed. That version still exists as
api.scentstudio.ai; the current platform is the second attempt, rebuilt on AWS
IoT Core with a containerised deploy path.
I documented it unusually thoroughly for a codebase with one author — an application blueprint, an integration note, design notes on device sourcing and deletion tracking. Partly because 3,619 commits is more than I can hold in my head. Mostly because a platform only one person can operate is a liability wearing the costume of an achievement.
What it turned into
AIRaware and LEAKaware are both built on this. Mould and water leaks are the same problem with a different molecule and a different customer: sensors in the field, readings that mean nothing raw, and a model that turns them into something worth acting on.
Two products from one platform, which is the entire argument for having built the platform carefully.