Staying useful when the real world changes.
The same smell can look different as humidity shifts, sensors age, airflow changes, or background air moves. The lab's job is to make readings reliable anyway.
What we research
Sensor drift is one of the biggest barriers in electronic noses. Aeralyte's long-term moat is calibration data: every controlled experiment and field deployment teaches the system how smell changes across environments.
The data flywheel
More experiments produce more calibration data; better drift correction produces more reliable fingerprints; more reliable fingerprints make deployments more valuable — which produces more experiments.
Humidity and temperature (SHT40) are tracked as first-class confounders, not afterthoughts, so the model can learn what to ignore.
How we keep results honest
Two checks are designed into the pipeline and will gate every bench and chamber result. Sensor-family ablation: remove one family and re-evaluate, so we know no single sensor is a crutch. Adversarial label-shuffle: scramble the labels and confirm accuracy collapses to chance — proof a model learns structure, not leaks.
The question we care about most — real drift across humidity, sensor aging, and background air — is exactly where data collection starts. Recording leads with the confounders: humidity-only controls and blanks sit alongside every exposure, so the boring air is on the record before any claim about the interesting air Bench. The rig build in flight carries these controls from bench toward chamber.
Further reading
- →Controlled sniffing — the sampling context we track against drift.
- →Data contracts — the calibration data the flywheel runs on.
- →Truth gates — the ablation and label-shuffle checks as gates.