At 7:14 on a Tuesday morning in November, a sensor in our neighbourhood reads AQI 412. The number is a fact. A neighbour, retiring after a walk that has left her short of breath, types into the JanVayu app: the air today tastes metallic, my chest is heavy, I won’t take the children out. That is also a fact.
Neither fact, alone, tells the school principal four streets away whether to keep the children indoors for assembly.
JanVayu has spent two years collecting both kinds of facts side by side. The sensor readings are dense and continuous. The testimonies arrive in bursts, in vernacular, with the kind of detail no instrument captures — what the air tastes like, whose lungs are loud, which window stays shut. The repository was always the easy part. The hard part was what to do with it next.
What the assistant does
A sensor on the left, the AI in the middle, a resident testimony on the right. The two streams feed in; the assistant sits in between.
The assistant takes a question from a resident — should we hold assembly outdoors at 7:30am?, is it safe for my asthmatic mother to walk to the temple this evening?, what is the air doing on the lane behind the school? — and reaches into the two streams at once. It pulls the sensor readings for the hour and the lane in question. It pulls the testimonies people have left within the same window. And it produces an answer in language the resident can act on: the readings are high and rising, residents nearby report the air as heavy, defer assembly to 8:30 or move indoors.
That is the whole thing. It does one job and refuses every other job.
What the design refuses
Most AI tools for environmental data make the opposite bet. They promise prediction. They promise dashboards. They promise to summarise a city’s air quality at country scale. The temptation is to build for the policy audience because the policy audience writes the cheques.
The JanVayu assistant refuses all of that. It does not predict tomorrow’s AQI — the existing models do that, and the prediction is rarely what a resident is asking for. It does not produce a city-level summary — the average of a city’s air is the place inequities go to hide. It does not score neighbourhoods against each other — that is a different project, and a more dangerous one.
What it does is sit between one number and one testimony, and translate. The translation is for one person at one moment. The technology is exactly as ambitious as that.
What this is an argument about
Measurement work in air quality has long had a quiet hierarchy. The sensor is the fact. The resident is the experience. Reports speak about reconciling the two and then mostly describe the sensor, with anecdote added for flavour.
The shape of the assistant on a phone: a short question goes up, a longer grounded answer comes back.
The assistant is a quiet argument against that hierarchy. The sensor reading and the testimony are both inputs, of equal standing. The reading is precise about parts per million. The testimony is precise about what the air tastes like and whose chest is loud. The two together produce something neither produces alone — a piece of advice an actual person can act on, on an actual morning, for actual children.
This is not a small technical claim. Most evaluation practice in this country still treats the sensor as the ground truth and the lived experience as confounding noise. JanVayu was built on the opposite assumption, and the AI assistant is the moment that assumption stops being a value statement and becomes an operating tool.
What is next
The assistant is live and learning. The corners we are watching are the ones where the two streams diverge — the morning the sensor says 220 but the neighbourhood says heavier, the afternoon the sensor says 90 but residents are reporting eye irritation in three lanes. Those are not errors. They are findings. Each divergence is a place where one of the two facts is reaching past the other, and a place where the design of public-interest air measurement needs to listen more carefully.
If you find a way the assistant is being asked the wrong question — or being asked a question it should refuse — write in. The tool gets sharper every time somebody pushes back on it.