The numbers Brief Episode map Takeaways Charts Quotes Clips Glossary Go next Transcript

Episode 146 · Deep tech · 36 min

Sell the model, not the monitor

Ambee began as a father's home-built air sensor in a Bangalore bedroom, and the number it returned — 600 to 700, against the 20-odd the nearest government station 14 km away was reporting — became the company. The founders' claim is that the hardware was never the product: climate data, fused from ground sensors, traffic feeds and NASA's hyperspectral satellites, is what pharma majors, hedge funds and ad platforms now pay for.

Climate & energy Data & digitisation AI & machine learning

Madhu and Akshay

Co-founders, Ambee · with Vishal Krishna

36:13

Said in this episode

▶ 2:20

600–700 vs 20–25

Bedroom sensor against the official station

Madhu's home-built monitor versus the reading from the nearest government station, 14 km away at Silk Board, on the same Sunday; the unit was never named on air.

▶ 10:45

99%

Share of air pollution that is human-made

Which is why the model ingests transport, buildings and industry rather than sensor readings alone — you have to measure the activity, not just the air.

▶ 13:33

3 hrs → 30 sec

NASA hyperspectral processing pipeline

The speed-up Ambee engineered on early-access satellite feeds, and part of why NASA scientists came on as advisers.

▶ 17:43

$600 → $250K

First pollen contract, then and now

A first contract at $600 a month with a breathing-products brand; Akshay says the same relationship is worth a quarter of a million dollars today, though the period for that figure was not stated.

▶ 20:53

3 weeks

Wildfire lead time in the LA back-test

Back-testing put the January Los Angeles fire areas in the medium-to-high-risk band by around 14–15 December; this was a back-test, not a live call.

▶ 25:02

38 of 52

Essential climate variables measured

The foundational data layer under Ambee's APIs, demand forecasts, wildfire risk tiles, GIS layer and health products.

The brief

The argument in sixty seconds

Ambee's claim is that the air over any given street is effectively unmeasured, and that the gap is a market. Madhu built a monitor because his six-month-old son was waking every night choking; the nearest government station, 14 km away, reported 20 to 25 after Sunday rain while his own sensor read 600 to 700, and three months of tracing led to a garment factory burning waste after the workers went home. Selling monitors — a couple of hundred, sourced from China — turned out to be a logistics and branding business, so they paid autorickshaw drivers ₹100 a week to carry sensors, then abandoned hardware entirely: data, Akshay argues, is by definition infinitely scalable. Because 99% of air pollution is anthropogenic, the model had to ingest transport, buildings and industry alongside satellite aerosol data, and the forecasting algorithm they published is what pulled NASA's hyperspectral programme towards them — where a three-hour processing job was rebuilt to run in under thirty seconds. The commercial twist is that India does not buy data, so the same physics was re-sold to the West as pollen: a first contract at $600 a month that is worth a quarter of a million dollars today, then pharma clinical trials and demand forecasting, and now wildfire risk, after a back-test put the January Los Angeles fire zones in the high-risk band three weeks early. Eight years in and on the edge of profitability, the bet is that owning 38 of the 52 essential climate variables is defensible in a way a trading business never is.

Worth your time if you are

Deep-tech founders choosing between devices and data

Climate and geospatial data scientists

B2B operators whose home market won't pay for the product

Investors sizing climate intelligence as a category

Parents who have never checked the air on their own street

Episode map

Where the conversation travels

Every block is a chapter, coloured by what it's about. Click any of it to jump straight to that minute on YouTube.

Climate & energyData & digitisationAI & machine learningFounder journeySales, GTM & growthHealthcare

01Cold open: a baby who couldn't breathe 0:00 An investor-podcast intro gives way to the origin story — Madhu moves from Hyderabad to Bangalore around 2015-16, his healthy six-month-old son starts waking every night coughing and choking, doctors call him asthmatic and say there is nothing more to do, and a doctor friend suggests the air. 02The nearest station was 14 km away 1:48 A Google search put the nearest monitoring station 14 km away at Silk Board reporting a benign 20 to 25 after Sunday rain, so he assembled a sensor in four or five hours, watched it read 600 to 700, and spent three or four months tracing the nightly spike to a garment factory burning something after the workers went home. 03Three founders and a problem statement 3:20 Door-to-door discovery with other parents, a startup leadership programme that introduced him to Jaideep and then to Akshay, a son who recovered in ten days without medicine once the family moved, and Ambee started in September 2017. 04Sensors on autorickshaws, monitors from China 4:42 A couple of hundred imported monitors in, Akshay says they were sinking into a logistics and branding game, so they paid autorickshaw drivers around ₹100 a week to carry small sensors and suddenly had street-by-street, hour-by-hour coverage — dense enough that a Google accelerator in Switzerland offered to buy the data if they covered three or four Indian cities. 05Data, by definition, is infinitely scalable 7:41 Unable to fund an operations-heavy rollout, they turned the sensors into training data for models instead, exploiting an India that supplies almost every climatic condition across Delhi, Bombay and Bangalore — and on 4 January 2019, at their first board meeting, the model results convinced Akshay this could scale. 0699% of air pollution is made by people 10:00 Reframed as a data-science problem, the model had to track human activity — transport, buildings and industry mapped through Google Places and historical traffic feeds — fused with European Space Agency columnar satellite data on aerosol optical depth, population density, topography and vegetation, yielding a forecasting algorithm with a lower error than standard interpolation, which they published. 07How the NASA door opened 12:47 The paper drew a referral into NASA's new hyperspectral satellite programme, where Ambee was already writing tooling against a slow FTP feed and rebuilt a three-hour processing job to run in under thirty seconds; three NASA scientists came on board and still advise the company. 08What early access actually buys 14:03 Madhu argues a model is only as good as the data you can feed it and that you need three summers and four or five winters to learn the interplay — access that carried Ambee from air quality into a worldwide pollen map classifying tree, weed and grass species, and then into wildfire forecasting. 09Nobody in India buys data 15:35 Angel money came in on the promise of selling monitors and a seed round landed just before COVID, but by late 2021 the models were accurate and the domestic market was not a data-driven economy — and pollution, the obvious export, is not a Western pain point. 10Every breathing brand on LinkedIn 16:53 Pollen behaves almost exactly like pollution and Bangalore's 150 to 200 years of imported trees gave them ground truth, so they cold-messaged heads of marketing at breathing-products brands with predictive pollen for ad placement; the first contract was $600 a month and the same relationship is worth a quarter of a million dollars today. 11Why pharma tests you for nine months 18:22 Unlike a free maps app you forgive for being wrong, a US pharma advertiser who calls a pollen surge wrongly faces the FDA, a falling stock price and a class action — so buyers test the data for eight or nine months, after which a $10,000-to-$30,000 advertising line grows into clinical trials and demand forecasting, because climate changes what people buy. 12Wildfire, found inside a broken model 20:20 Physically impossible air-quality spikes in the wilderness looked like model failure until back-testing revealed nearby wildfires; the resulting model, run backwards against the January Los Angeles fires, had put the medium-to-high-risk areas in the band around three weeks earlier, and is sold as a weeks-to-months risk window rather than a time-and-place prediction. 13Climate intelligence, and who trusts a forecast 22:30 Madhu positions climate intelligence as a layer settling on top of business intelligence in a world of rising disasters and allergy incidents, and argues trust is earned product by product — the analogy is a new Apple device inheriting the brand — by never undercutting the science for the sake of a use case. 1438 of 52 variables, then the stack 24:31 The foundational layer is data — 38 of the 52 essential climate variables at high granularity — with APIs above it, and then demand forecasting, clinical trials, wildfire risk, illness and flu forecasts, map tiles and a GIS layer packaged into separate products per use case. 15The other marriage 25:47 Asked about eight years of co-founding, Akshay notes he has been married less time than he has been at Ambee, describes equal equity — Madhu holds three shares more, roughly 0.1% — and a deep-tech company perpetually six or eight months from ruin, while Madhu recalls the years without salaries and the debt taken on so no employee was ever paid late; moving one founder to the US turned a wary customer into a hockey stick within a month. 16On the edge of profitability 32:00 Profitable last year before they handed out big hikes, and deliberately not tracking profit as the north-star metric, they now field inbound from hedge funds, sovereign and mortgage funds and top-three global pharma companies who ran diligence and found a name brand — while rivals, including one co-founded by a Nobel laureate, are treated partly as a future supply of satellite data to consume.

Takeaways

Ideas to carry out of this hour

The measurement gap was the entire business

The official picture said the air was fine: the nearest monitoring station was 14 km away at Silk Board, and after Sunday rain it read 20 to 25. The sensor Madhu built in an afternoon, in the room where his son was choking every night, read 600 to 700. Three or four months of walking the neighbourhood found the cause — a garment factory burning waste at night, when the workers had gone — and established the pattern that became the product: air changes from morning to night, from one end of a road to the other, before and after rain, and almost none of it is measured.

▶ 1:48 ▶ 2:20 ▶ 2:49

Hardware was a trap; the data was the asset

They sourced monitors from China and sold a couple of hundred, made a little money, and concluded quickly that they were in a logistics and branding game rather than building anything defensible. The bridge was cheap: small sensors carried by autorickshaw drivers for about ₹100 a week, an app to collect the feed, and street-level density no one else had. But scaling that operationally meant salaries they could not pay, so they inverted it — the sensors became training data, because data, unlike a fleet, is infinitely scalable if the models are right.

▶ 5:42 ▶ 6:27 ▶ 7:41

To model air, model people — then earn the satellite

Ninety-nine percent of air pollution is anthropogenic, so sensors alone were never going to be enough: the model had to follow energy consumption into emissions, mapping transport, buildings and industry through Google Places and historical traffic feeds, then fusing that with European Space Agency columnar data on aerosol optical depth, population, topography and vegetation. Standard interpolation methods fell short, so they wrote their own forecasting algorithm, got the lowest error, and published it. The paper is what produced the referral into NASA's hyperspectral programme — not a pitch.

▶ 10:45 ▶ 11:45 ▶ 12:31

Early access is won by contributing, not asking

NASA was launching the first hyperspectral satellite and offering ground-station access to early adopters. Ambee was already writing code against the simulated data sets, working around a slow FTP delivery mechanism, and rebuilt a pipeline that took three hours into one that ran in under thirty seconds. Three NASA scientists came on board and still advise the company. Madhu's argument for why that access compounds: a model is only as good as the data you feed it, and you need three summers and four or five winters before the variables reveal their interplay.

▶ 12:47 ▶ 13:18 ▶ 14:18

India wouldn't buy the data, so the physics was re-sold as pollen

By late 2021 the models were accurate and the market was missing: India is not a data-driven economy, and the natural export markets in the EU and US do not have India's pollution problem. Pollen was the arbitrage — it behaves almost identically to air pollution, and Bangalore, with 150 to 200 years of trees imported from across the world, gave them first-hand ground truth. The go-to-market was equally unglamorous: cold LinkedIn messages to heads of marketing at breathing-products brands, pitching predictive pollen so ads could be placed where people would be sneezing in two or three weeks. The first contract was $600 a month; the same relationship is a quarter of a million dollars today.

▶ 16:21 ▶ 16:53 ▶ 17:38

Pharma's paranoia is the moat, once you survive it

If Google Maps is wrong you apologise for being late and keep using it. If a US pharma brand advertises a pollen surge that does not happen, the FDA opens an inquiry, the stock drops and a class action follows — so buyers test a data vendor for eight or nine months before signing anything. That test is the barrier to entry: once through, a $10,000-to-$30,000 advertising line grows into clinical trials, where Ambee says it is the only allergy data provider, and into demand forecasting for a country where even two-day delivery cannot fix stock stranded in the wrong state.

▶ 18:37 ▶ 19:00 ▶ 19:21

The wildfire product came out of a model that looked broken

Two years before this conversation, air-quality readings started spiking to physically impossible values in remote locations — the same model that worked everywhere else appeared to be failing. Back-testing showed the spikes were real: a nearby wildfire changes air quality almost instantly, the way an incense stick fills a room. That accident became a forecasting product, and back-tested against the January Los Angeles fires it had flagged the medium-to-high-risk areas around 14 or 15 December, roughly three weeks ahead. Crucially it is sold as a risk window — take precautions in these areas over the coming weeks — not as a claim about tomorrow at 3 p.m.

▶ 21:09 ▶ 20:53 ▶ 21:55

The numbers, drawn

What the episode measures

Every figure below was said on air — timestamps included, caveats kept.

Conversation share

portion of the hour spent on each theme

Climate & energy22%

Data & digitisation18%

AI & machine learning14%

Founder journey12%

Sales, GTM & growth11%

Healthcare8%

Computed from the chapter map of this episode.

What the nearest station missed

air-quality reading as stated on air

As stated in conversation: the official reading after Sunday rain 'looked 20, 25', while the home-built sensor 'was showing 600, 700'. Midpoints of the two ranges are plotted; the measurement unit was never specified on air.▶ 2:20

The foundational data layer

essential climate variables

Measured by Ambee73%

Not claimed27%

As stated in conversation: 38 of the 52 essential climate variables are measured at high spatial granularity. Nothing was said on air about the remaining 14.▶ 25:02

Worth keeping

Lines that stay

The nearest air monitoring station is 14 km away, in Silk Board. It had rained in Bangalore and it looked 20, 25 — everything good about it. The air quality from my sensor was showing 600, 700.

— Madhu ▶ 2:20

We said we should build a data-first model, because data by definition is infinitely scalable — if you can build the right models.

— Akshay ▶ 7:41

This is the sophistication of our sales process: we found every breathing brand on LinkedIn and messaged their head of marketing. Our first contract was $600 a month. That relationship is a quarter of a million dollars today, with the same company.

— Akshay ▶ 17:43

A deep-tech company is always six or eight months away from ruin, so we can't afford to focus on anything else. Our egos played less of a role than our need to get things done.

— Akshay ▶ 28:35

If we had started ten years before, there would be no competition. Ten years later would probably be too late. But today we have very little competition.

— Madhu ▶ 35:10

Clips that travel

Short on time? Start here

Parents who have never checked the air on their own street

The night the sensor read 700

The four-hour build, the station 14 km away calling it fine, the garment factory burning waste at night, and a son who recovered in ten days without medicine.

1:33 → 4:42 · 3 min ▶ Watch clip

Deep-tech founders choosing between devices and data

Why they walked away from hardware

The logistics-and-branding trap, autorickshaws carrying sensors for ₹100 a week, and Google's Switzerland offer that turned data into a business model.

4:42 → 8:55 · 4 min ▶ Watch clip

Climate and geospatial data scientists

Modelling people, and the NASA door

Fusing ground sensors with satellite aerosol data, the published algorithm that earned the referral, and a three-hour pipeline rebuilt to thirty seconds.

11:45 → 15:35 · 4 min ▶ Watch clip

B2B founders whose home market won't pay

Nobody in India buys data

The pollen arbitrage, LinkedIn cold outreach, $600 a month becoming a quarter of a million, and why pharma tests a vendor for nine months first.

16:21 → 20:20 · 4 min ▶ Watch clip

Co-founders negotiating equity, roles and ego

Eight years, three shares apart

Equal equity by choice, debt taken on so nobody was paid late, and the founder relocation that turned a customer's doubt into a hockey stick.

26:33 → 32:00 · 5 min ▶ Watch clip

Glossary

The jargon, unpacked

Remote sensing / Earth observation

Measuring the atmosphere and land surface from satellites or aircraft rather than from instruments on the ground — the discipline Madhu stumbled into while trying to explain where pollution was coming from.

Hyperspectral imaging

Capturing hundreds of narrow bands of light instead of a few broad ones, so a satellite can tell materials apart — smoke, aerosols, one plant species from another. Ambee was an early adopter of NASA's hyperspectral feed.

Aerosol optical depth

A satellite measure of how much light airborne particles block through a vertical column of atmosphere; the columnar, low-resolution signal Ambee fuses with ground sensors to estimate what people at street level are actually breathing.

Anthropogenic

Caused by human activity. The founders' working figure is that 99% of air pollution is anthropogenic, which is why the model tracks traffic, buildings and industry rather than air alone.

Interpolation

Estimating a value where you have no sensor from the handful of places where you do. Standard interpolation methods fell short for air quality, so Ambee published its own lower-error forecasting algorithm.

Essential climate variables

The standardised set of physical, chemical and biological measures used to track the climate system; Ambee says it measures 38 of the 52 at high spatial density.

Climate intelligence

The category the founders describe as forming on top of business intelligence: climate and environmental forecasts fed directly into commercial decisions — ad placement, drug stocking, clinical-trial design, insurance and infrastructure risk.

Connections

If this resonated, go here next

EP 30 · Deep tech

Full transcript

The whole conversation, searchable

Auto-generated captions, lightly cleaned. Click a timestamp to open that moment on YouTube.