Galli Glow measured what children in ten Mumbai classrooms were actually breathing — PM2.5, PM10, CO₂, heat and humidity — then built a model that predicts it from satellite data alone, for schools that will never afford a sensor.
01 — The problem
Boston Public Schools spent seven million dollars and hired Harvard to find out what its children were breathing. A Mumbai municipal school has no such budget, no classroom CO₂ standard to be held to, and no data at all. The first job was not to raise an alarm — it was to find out.
Particles small enough to cross from lung into blood. Here they come from traffic, construction and chalk dust, which alone can raise a poorly ventilated classroom by almost 400% (Ravindra et al., PMC11483887).
Not poisonous at these levels — it is a direct measure of how stale the air is. Past 1,000 ppm, measurable falls in concentration and decision-making show up in the research.
In a room with no cooling, indoor temperature just tracks outdoor temperature. The roof is the biggest lever — a metal sheet radiates heat straight down onto the children beneath it.
The coarser fraction — road dust, construction, chalk. It settles faster than PM2.5 and is driven more by what happens inside the room than by what drifts in from outside.
Mirrors the outdoor air, modulated only by airflow. Through the monsoon Mumbai sits near saturation, and above 80% you get mould, damp and discomfort.
The monsoon baseline, measured against published guidance
From 15 June to 30 August all ten stations rated Healthy: PM2.5 11–16 µg/m³, PM10 20–28, CO₂ 440–500 ppm. The whole campaign fell inside the monsoon, when rain scavenges particulates from the city's air. Run the same pipeline on winter inputs and several classrooms would be expected to move into Moderate or High-risk — that is the next step, not a footnote.
Boston has alerted at 1,000 ppm for years. India has no classroom-specific CO₂ standard at all — so even where a room is stale, there is nothing on paper a school can be held to, and no threshold a headmaster can point at.
Harvard studied Boston. SchoolAIR studied Portugal. Delhi has vendor case studies with no published outcomes. Mumbai has thousands of classrooms and no budget to instrument them — so the useful question is whether a room's air can be predicted without putting a sensor in it.
02 — How it started
Five moves, each one caused by the failure of the last. This is the shape of the project — not a plan that was drawn up in advance.
Years of volunteering at Vatsalya led to teaching spoken English at BPES, an underserved Marathi-medium school in Bandra. Spending time there, she stopped noticing the lesson and started noticing the room: hot, stuffy, dusty. As someone with asthma, she knew what that meant.
"I wasn't looking for an air-quality problem."
First came the things you can point at. A GPS-tagged Galli-Map audit ranked every classroom by need in June 2026. In July, solar panels donated through an internship at Vivid Green became 16 lighting and fan units across BPES and Vatsalya, with volunteers, teachers and students trained to maintain them. The Shwas fundraiser then raised ₹120,000 from 20 donors for the sensors.
The air was the part nobody had measured.
Five separate readings still don't answer the only question a teacher actually asks: is our classroom OK today? So she built the HCEI — one 0–100 score, weighted across all five and read against published guidance.
Five readings, one number, on fixed weights.
Sensors in every classroom don't scale to schools with no budget. So from September 2026 the logged data trained a model that predicts classroom conditions from satellite and reanalysis products alone — aerosol depth, trace gases, weather, vegetation, population — with no hardware in the room at all.
Validated on a classroom it had never seen.
Predicting a problem doesn't solve it. CO₂, PM2.5, heat and humidity become characters the children name and design themselves; the game's choices become Clean Air Missions they run in their own classroom, and the HCEI tells them whether it worked.
"The classroom becomes the final level of the game."
03 — The instrument
A 0–100 score per classroom, built from five logged parameters. Each is scored continuously against published guidance, then combined on fixed weights. One number a teacher can read at a glance, with the parameter that earned it sitting underneath.
Scored continuously, not banded. Each pollutant runs linearly from a good guideline value down to a poor one and is clipped at both ends; temperature and humidity take full marks inside a comfort band and taper outside it. The weights sum to one. A room slightly over guidance therefore reads differently from one far over it — which is what makes the index usable for ranking classrooms rather than just flagging them.
Try it — the sliders open on the measured baseline
04 — The dashboard
During Step 1 this is a monitoring tool. During Step 2 it becomes part of the intervention itself — the children see the traffic light, the administrator sees what to fix and what it costs. Same instrument, two audiences.
One number, 0 to 100
Higher is healthier. At or above 80 a room rates Healthy, 50–79 Moderate, below 50 High-risk. The weakest of the five parameters sits underneath, because that is what the score is really telling you.
Which of the five is weakest
Each tile carries the reading, the guideline it is judged against, its weight in the index, and the shape of the school day. One weak parameter is enough to pull the whole score down.
Measured against guidance
Each reading set beside the guideline it is judged against. Through the monsoon every station sat inside guidance. That the index returns Healthy rather than defaulting to alarm is a property worth having in a screening tool.
What to fix, and for how much
It names the weakest parameter and returns three costed fixes for it, cheapest first, drawn from a library where every entry carries a rupee figure and a published source.
Ten classrooms · 230 students, ages 6–16 June–August 2026
Weakest parameter: RH
scoring 75 of 100
| Time | PM2.5 µg/m³ | PM10 µg/m³ | CO₂ ppm | Temp °C | RH % | HCEI |
|---|
The engine finds the parameter with the largest weighted shortfall against guidance, pulls the measures that act on it, and ranks them by cost, then by installation difficulty, then by strength of published evidence. It does not attempt to separate what came in from outside from what was made in the room — that would need an outdoor sensor at each site, which this deployment does not have.
05 — What actually works
BPES cannot afford HVAC or commercial filtration. Everything in this library is costed in rupees, sourced to published evidence, and ordered so that the free ones come first. Filter it the way a headmaster would: by what's broken, and by what's in the drawer.
Showing 12 of 34 measures · 11 cost nothing at all
06 — Where
Ten schools, most of them serving communities with limited resources, deliberately spread across twenty kilometers of the city — from the southern island districts through Bandra and Worli to Chembur. Testing the model across genuinely different surroundings is the only way to know it generalizes.
Named in the paper: Saboo Siddiqui School and Saboo Siddiqui Junior College — co-located, sharing one satellite pixel and kept as separate rooms — along with Mahatma Gandhi School, Maratha Mandir, Bandra and Vatsalya. Hourly logging from 15 June to 30 August 2026.
Marathi-medium school founded in 1954, serving underserved communities in Bandra. Tvishaa taught spoken English here, ran the Galli-Map audit across its twelve classrooms, and installed the first solar lighting and fans before any sensor went in.
A Mumbai children's home Tvishaa has been part of since she was young — first celebrating birthdays there, later teaching English. Vatsalya opened the door to BPES, supplied the volunteers trained to maintain the solar units, and carries one of the ten monitoring stations.
07 — Who builds it
The sensors are cheap and the pipeline is reproducible on purpose — a method that claims to scale to unmonitored classrooms has to work with hardware those schools can actually afford, and be auditable by someone who wasn't there.
Galli Glow's model for predicting classroom air quality is being developed under his guidance. The baseline study is complete and the model has been validated against both held-out classrooms and held-out weeks.
One inexpensive sensor per classroom, logging five parameters about once an hour at irregular moments. Offline entries were treated as missing and the rest averaged into clean hourly values. The hardware is modest on purpose — a method that only works with research-grade instruments is no use to the schools this is for.
Satellite and reanalysis data predict what a classroom's air is doing without a sensor in it. Validated leave-one-classroom-out — trained on nine rooms, asked about a tenth it has never seen — and again on the final weeks it was never trained on. R² 0.62–0.76, RMSE 1.1–1.5 µg/m³ at every site.
Built on
SHAP puts month first, then the satellite pollution layers — CAMS particulates and aerosol optical depth — with vegetation, elevation and urban form behind them. School hours and day of week matter least. That ordering is the finding: indoor PM2.5 is governed by outdoor air infiltrating the room, not by what happens inside it.
Two safeguards stop the model cheating. Classroom identity is excluded as an input — a model told which room a reading came from could memorise it and then fail at a new school, which is the whole problem. And indoor CO₂ never informs the particulate models, because a school with no sensors would have neither reading.
08 — Founder
Tvishaa founded Galli Glow in 2025 to close the infrastructure gap in Mumbai's underserved schools — first with solar installations and renewable-energy workshops, then with the thing nobody was measuring at all. She is in Grade 11 of the IB Diploma Programme at Dhirubhai Ambani International School, on the robotics team, and has asthma. That last fact is why she noticed the classroom.
The work is deliberately unglamorous. Consent letters before sensors. A real NDIR CO₂ module instead of a cheap one that would have passed unnoticed. Leave-one-school-out validation because a model trained on three buildings cannot honestly claim to predict a hundred. She would rather publish a smaller finding that survives scrutiny than a larger one that doesn't.
Outside the lab she is a trained aerialist with Fly High and a theater practitioner who has performed at the NCPA — which is less of a detour than it sounds. The whole Step 2 intervention rests on whether four invisible gases can be made into characters a six-year-old will argue with.
The paper
Tvishaa Prabhu · Dhirubhai Ambani International School, Mumbai · submitted to IJSDR
Ten pilot schools, hourly measurement from 15 June to 30 August 2026, gradient-boosted prediction from satellite and reanalysis predictors, validated leave-one-classroom-out and forward in time. R² 0.62–0.76, RMSE 1.1–1.5 µg/m³.
"What stood out from the beginning was her effort to understand what would genuinely be useful and accessible to the children, rather than developing technology in isolation. The children will participate in testing and shaping the game rather than simply being recipients of the project." — Bombay Presidency Education Society
"We have seen her grow from participating in activities with our children to independently using her skills to address challenges affecting underserved communities. Her experiences here have shaped her understanding that meaningful solutions need to be developed with the communities they are intended to serve." — Vatsalya Foundation
09 — What was done
Light and fans first, because those were the problems anyone could name. The air was the part nobody had measured — and measuring every room was never going to be affordable, which is what made a model the point rather than a flourish.
Trained on nine schools and asked about the tenth, the model reproduces daily indoor PM2.5 with R² between 0.62 and 0.76 and RMSE of 1.1–1.5 µg/m³ — at every site, across very different parts of the city. Feature attribution puts satellite pollution layers near the top and occupancy near the bottom, confirming the outdoor-to-indoor link rather than assuming it.
Every station rated Healthy, and that is a real result rather than a null one — the index reflects genuinely good conditions instead of defaulting to alarm. But the campaign sat entirely inside the monsoon, when rain scavenges particulates from the city's air. Winter inputs would be expected to push several classrooms into the Moderate or High-risk tiers.