The sensor measures. The AI makes it honest.

The detection pipeline

Four models between your breath and your number.

Model 01 — Breath sample validation

The single biggest source of garbage readings on consumer breathalyzers is the sample itself: a blow that is too short, too weak, interrupted, or contaminated by mouth alcohol from a recent sip. The KeyCheck and Pro kits record the full exhalation flow profile — pressure and alcohol signal over the seconds you blow — and a small on-device classifier checks its shape against thousands of labelled good and bad samples.

• Short or weak blows produce a truncated curve that never reaches deep-lung (alveolar) air, which is the only air that correlates with blood alcohol. The model rejects these and coaches you to blow longer.
• Mouth-alcohol contamination — testing minutes after a sip — shows up as a sharp early spike that decays abnormally fast. Instead of reporting a terrifying (and false) 0.20, the app flags it and starts a 15-minute retest timer.
• Interrupted samples (coughs, laughs, second breaths) create discontinuities the classifier catches, prompting a clean retry.

Model 02 — Environmental & drift compensation

Semiconductor and fuel-cell sensors are physics, and physics has opinions about the weather. Cold air, dry air, altitude and simple sensor aging all bend the raw signal. Each device carries a compensation model trained on paired readings across environmental conditions; it takes the raw voltage plus onboard temperature and humidity readings and outputs a corrected estimate.

The same model tracks drift: as a sensor ages, its baseline shifts in a predictable pattern. The Pro app watches your device's baseline between uses, silently corrects early-stage drift, and tells you when correction is no longer enough — that is what triggers the calibration reminder, rather than a dumb one-year timer.

Model 03 — Your personal elimination model

The textbook says BAC falls about 0.015% per hour. Your body did not read the textbook. Actual clearance varies with body composition, liver enzymes, medications, sleep and more. When you log serial readings in the Pro app — say 0.09 at midnight, 0.06 at 2 AM — the app fits a personal elimination curve to your data.

That powers the sober-clock: a live prediction of when your BAC will approach zero, with a confidence band that starts wide and tightens as more of your readings accumulate. Two design rules are hard-coded:

• The clock always displays the conservative edge of the confidence band — the latest plausible time, never the earliest.
• It predicts when you will be near zero, not when you will slip under 0.08 — because the legal limit is a prosecution threshold, not a safety target.

Model 04 — Anomaly detection

Human BAC follows physiological rules: it cannot triple in ten minutes, and it cannot fall off a cliff. A lightweight anomaly detector compares every new reading against your recent trajectory and those rules. Implausible values — a defective strip, a contaminated sample, a sensor glitch — get flagged with a retest prompt instead of being presented as truth. One bad number believed at 1 AM is exactly the failure mode this exists to prevent.

On your device, and only yours

Every model in the stack is small enough to run on-device — in the KeyCheck's microcontroller or your phone for the Pro app. No cloud round-trip, no account required, no reading ever leaves your hardware. Your elimination profile is arguably health data, and we treat it that way: local by default, deletable in one tap, never sold, never shared.

What the AI will never do

• It will not turn a consumer device into an evidential instrument — police-grade equipment and blood draws remain the only measurements courts accept.
• It will not tell you it is safe to drive. Every output rounds toward caution, and every screen that shows a number also shows the same sentence: any reading above zero means do not drive.
• It will not diagnose impairment from other substances, fatigue or medication — alcohol is the only thing these sensors measure.

Engineering details describe the OWI Labs detection stack at a general level; general information, not medical or legal advice.