How VANTIL decides what to believe.
A monitor that raises too many false alarms gets ignored. VANTIL is built in layers, each one answering a different question before the next one is allowed to speak.
- 1
Can this reading be trusted?
Every reading is checked against what the sensor can physically produce and what its neighbours say. A failing or warming-up sensor is marked as such, and its readings are kept out of the analysis instead of being quietly averaged in.
- 2
Is it within a normal range?
Readings are compared with published reference ranges for particles, carbon dioxide and comfort, so the system always has an absolute yardstick.
- 3
How much does it matter?
A health gradient grades readings from comfortable to concerning, so a small excursion and a serious one are not described in the same words.
- 4
Is it normal for this home, at this hour?
The system learns a baseline for each hour of the day in each home. Dinner time looks different from 3 a.m., and a home beside a freeway looks different from one in the hills. A change is judged against your own normal.
- 5
What changed, and what else moved?
When a reading departs from its baseline, the system groups it into one event, measures how large it was, and looks at what the other sensors, and the outdoor node, were doing at the same time.
Principles
Noticed is not the same as proven.
Say what is known
A pattern becomes a finding only after it has been tested on data the system had not seen. Until then it is labelled as an observation.
Say what is not
When the evidence is thin, the answer is "not sure yet", not a confident guess.
Count the mistakes
False alarms are tracked, reviewed by a person, and used to tighten the system. We do not tune the checks while an evaluation is running.
Keep people in the loop
Human feedback on what an event really was is how the system learns which patterns matter.
Limits
What VANTIL cannot do.
- It cannot name a specific chemical. It reports an index that responds to a family of compounds.
- It is built from low-cost sensors, so it is better at showing change than at precise absolute values.
- It does not diagnose, treat or predict any health condition. It is not a medical device.
- It is a prototype running in one home. How it behaves elsewhere is still to be learned.