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

Episode 156 · Deep tech · 40 min

The question that outlived the prize

At the dinner for his 2004 Bhatnagar prize, the politicians and administrators asked the laser spectroscopist one thing: what is the use for us? Two decades later the answer is a reader that fires light through sealed boxes and, because it is trained on what makes a chemical burn or blast rather than on a library of fingerprints, calls a never-before-seen compound suspicious. It was tested over three months on checked baggage at Bangalore airport.

Deep tech & hardware Founder journey AI & machine learning

Siva Umapathy

Founder, Gyrophint · with Vishal Krishna

40:03

Said in this episode

▶ 24:07

3 months

Trial at Bangalore airport

Run at level four on checked-in baggage, alongside the X-ray screeners who flag bottles and packets for questioning.

▶ 27:00

<5 sec

Time to scan a sealed box

The pitch for border and port use: bring the bag over, scan the box, and either raise the alarm or wave it through.

▶ 22:39

360°

Scattered light collected

Against the conventional 90 or 180 degrees, which he says wastes 250 to 280 degrees of the signal, and is the basis of one of the company's patents.

▶ 29:00

90–95%

Sepsis model accuracy

Varies with the sample, with false negatives put at under 3 per cent, on about 2,000 hospital samples of which only 200-250 were confirmed sepsis.

▶ 11:27

Investor presentations before yes

His own count over roughly a year and a half in 2021-22; convincing someone to invest, he says, is the hardest part of becoming an entrepreneur.

▶ 31:34

6–8 hrs

Battery on the backpack unit

Design complete and components being sourced, aimed at narcotics-bureau teams scanning on trains, buses and streets.

The brief

The argument in sixty seconds

Umapathy's claim is that contraband detection fails on epistemology, not optics: deployed scanners match a measured spectrum against a library, so a freshly cooked explosive or a new fentanyl analogue simply is not in the book and walks through. Gyrophint trains AI on the physical and chemical properties behind the response function, the burn and the blast, so an unknown compound comes back flagged suspicious, with the family it resembles shown on the detail screen. Two patented moves make the signal good enough to bet on: mathematics that strips Rayleigh scatter and stray light out of a Raman measurement, and collecting scattered light across the full 360 degrees rather than the customary 90 or 180, which lifts signal-to-noise enough that you never have to burn a biological sample with more laser. The proof he offers is unglamorous and specific: three months at Bangalore airport on level-four checked baggage, where the instrument named thinner in a painter's unlabelled bottle and camphor in a balm in under five seconds, and where the airport wrote to the ministry for trial permission and sent its screeners into the lab to say where the buttons should go. The same physics reads body fluids, at 90 to 95 per cent accuracy on sepsis with false negatives under three per cent, but he is holding that back six to eight months to consolidate contraband first. What thirty years of citations could not buy him was capital: forty presentations across a year and a half, in a country he says has gone 75 years without a single technology it can sell to the world.

Worth your time if you are

Academics weighing whether to spin a lab out

Deep-tech founders raising a first institutional round

Aviation and border-security buyers evaluating screening tools

Diagnostics teams building models on small clinical datasets

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.

Deep tech & hardwareFounder journeyAI & machine learningHealthcareProduct strategyLeadership & org

01A phone call from Oxford to IISc 0:00 Naganand opens Ideaspring's founder-in-focus format, and the professor retraces the 1988-89 call that reached him at Rutherford labs and brought him back to build a time-resolved laser spectroscopy group nobody else in India was running, built on two pulses, one to start a chemical reaction and one to watch it move. 02What is the use for us? 2:43 At the prize dinner the administrators skipped the citation and asked what the work costs and what use it is to them, a question that followed him onto the flight home and pushed an expensively funded basic scientist toward biology and biomedical signals. 03Can I see what is inside that bag? 4:14 Chairing a Delhi committee on explosives and narcotics detection in 2007-08 produced the founding question about abandoned baggage, and rewriting the Monte Carlo treatment for frequency-shifted light showed that a Raman signal could survive a box or a bottle. 04Two instruments, a patent, a banker brother 6:08 Instruments delivered to DRDO labs in Pune and Bangalore, a sepsis patent and colleagues who had already crossed from lab to company turned into a registered startup, with a brother newly retired from a bank taking the GST, PF and income-tax side he freely admits he does not understand. 05Forty pitches and a patience problem 8:15 His two fears were the rules for running a company and his own inability to convince investors, and he argues Indian deep-tech capital lacks the five-to-ten-year patience the West extends. It took a year and a half and roughly forty presentations before the pitch learned to speak about money rather than science. 06The age question, asked to his face 11:59 Naganand puts the uncomfortable VC thesis directly to a founder who has just retired from IISc (enthusiasm, sixty-hour weeks, changed habits, accumulated baggage) and says he decided to bucket this one differently. 07From being the boss to reporting 13:31 Waking at four, defending himself in meetings instead of directing them, holding his team and himself to a weekly-and-monthly review, shutting down a genuinely interesting research thread because it is somebody else's money, and studying go-to-market on YouTube while his wife wonders what happened to him. 08Why MRI cannot see chemistry 16:34 The original ambition was non-invasive chemical imaging of the body: cancer cells change composition as well as shape, but MRI, functional MRI and X-ray all read density, which is why disease still has to be confirmed by drawing blood. 09A spectrum treated as an image 18:40 From 2014-15 he had an MTech computer-science hire treat the scattering signal exactly like an RGB image, frequencies and relative amplitudes in a three-dimensional hyperspectral picture, and two to three years later the data was good enough to publish among the world's first AI-in-spectroscopy papers. 10Suspicious beats matched 20:16 Because the model learns the properties that produce an explosion or a burn rather than a fingerprint library, a new explosive or a new fentanyl family returns as suspicious with its nearest relatives listed. That, he argues, is the one thing separating Gyrophint from everyone else doing this. 11360 degrees of scattered light 21:50 Two differentiators he traces to twenty-five years of papers and patents: signal analysis that delineates Rayleigh scatter, background and stray light from the real signal, and collecting the full 360 degrees of scattering instead of wasting 250-280 of them, which is how you get signal off a biological sample without burning it. 12Three months at Bangalore airport 23:54 A level-four checked-baggage trial turned the screener-versus-passenger argument into an answer: a painter's white bottle of water-like liquid read as flammable until he admitted it was thinner, and a suspicious balm resolved to camphor, leaving the officer a call to make rather than a hunch. 13Borders, ports and a five-second scan 26:15 He sizes the market past airports to shipping ports and land borders where children and old women carry packets nobody can question, then credits Bangalore airport for writing to the ministry for trial approval and lending its screeners to the lab as design partners. 14The sepsis model and the kidney case 28:45 About 2,000 hospital samples yielded only 200-250 confirmed sepsis cases because confirmation itself takes three or four days; the model runs at 90-95 per cent with under 3 per cent false negatives, and once flagged sepsis in a patient the doctors had cleared, correctly as it turned out, via kidney failure. 15Backpack now, handheld next 30:48 Contraband is the ready product and sepsis slips six to eight months while more hospital samples accumulate; a battery-powered backpack unit is a month and a half out, a handheld for VIP movements and the narcotics bureau two and a half months behind it. 16India first, then the two-thirds 32:34 He wants six months to a year of Indian field use and a few hundred units before any international push, has the certifications and a Hindi interface done with Kannada to follow, and ends on the vision of an Indian instrument cheap enough for Africa and South America, and more professors turning papers into products.

Timeline

The arc, briefly

From a phone call at the Rutherford labs to an instrument on an airport baggage line.

1988-89A call reaches him at the Rutherford labs. A dean at the Indian Institute of Science asks whether he wants to come back, and he builds a time-resolved laser spectroscopy group nobody else in India was running: two pulses, one to start a chemical reaction and one to watch it move. Funding in the 1990s, he says, was never the problem.

2004The Bhatnagar prize, and the question at the dinner. The politicians and administrators skip the citation on chemical dynamics and ask what it costs and what use it is to them. A stranger on the flight home asks the same thing, and it pushes an expensively funded basic scientist toward biology.

2007-08He chairs a Delhi committee on explosives and narcotics detection. The founding question is an abandoned bag. Rewriting the Monte Carlo treatment for frequency-shifted light showed a Raman signal could survive a box or a bottle, and instruments followed to DRDO labs in Pune and Bangalore.

2014-15An MTech computer scientist is told to treat a spectrum as an image. Frequencies and relative amplitudes read the way RGB is read, a three-dimensional hyperspectral picture rather than a curve to be matched against a library.

2017-18The data finally gets good enough to publish. Two or three years on they are among the first in the world publishing AI in spectroscopy, and the model can be trained on the property that produces an explosion or a burn instead of on a fingerprint. A new fentanyl analogue comes back suspicious, with the family it resembles named.

2021-22Forty presentations, and nobody convinced. A year and a half of pitching taught him the problem was never the science: investors needed to hear how it turns into money, not why it is pathbreaking. His own indictment: 75 years, several Nobel laureates from this country, and not one technology India can sell to the world.

TodayThree months at Bangalore airport. On level-four checked baggage it named thinner in a painter's unlabelled bottle of water-like liquid, and camphor in a balm that had looked suspicious, each inside five seconds, leaving the screener a call to make rather than a hunch. The airport wrote to the ministry for trial permission and sent its screeners into the lab to say where the buttons should go. Contraband ships first; sepsis slips six to eight months while more hospital samples accumulate.

Takeaways

Ideas to carry out of this hour

A dinner-table question redirected twenty years of research

The Bhatnagar citation on chemical dynamics meant nothing to the politicians and administrators at the prize dinner; what they asked was what it costs and what use it is to them, and a stranger on the flight home asked the same thing. He was, by his own account, a man who had received a great deal of public money for lasers and basic science. The discomfort, rather than any market study, is what sent him into the biology division at IISc and, eventually, into a company.

▶ 2:58 ▶ 3:28 ▶ 3:43

Fingerprint matching guarantees you miss the new molecule

Every deployed system, he argues, compares a measurement against a library and needs an exact match, which means a novel explosive or a fresh fentanyl analogue is simply absent from the reference set. Gyrophint instead trains the model on the physical and chemical properties that produce the response function, the ability to explode or burn, so an unrecognised compound returns as suspicious and the detail screen names the family it resembles. Without that, he says flatly, they would be the same as the rest of the world.

▶ 20:16 ▶ 20:49 ▶ 21:28

The patent is in the 270 degrees everyone throws away

Raman scattering happens in all directions, yet instruments conventionally collect at 90 or 180 degrees and discard 250 to 280 degrees of the light. Collecting across the full 360, plus multiple excitations and multiple collections, lifts signal-to-noise where it matters most, on biological samples, where the obvious fix of turning the laser up simply burns the specimen. Alongside that sits the older lab skill: mathematics that separates Rayleigh scatter, background and stray light from the signal, out of papers he says are still cited and used.

▶ 22:23 ▶ 22:39 ▶ 23:10

Treat the airport as a design partner, not a customer

Bangalore airport's leadership wrote to the ministry seeking approval to trial the technology, then ran it for three months on level-four checked baggage, and the civil aviation security office helped shape the trial design. More valuable still, the airport sent working screeners into the lab to specify the instrument: probe shape, button placement, what the software should offer. His verdict is that there is nothing like building a technology with the users telling you what they want, and Naganand's point is that India needs far more institutions willing to front-end a trial like that.

▶ 24:22 ▶ 27:44 ▶ 28:14

In diagnostics the labels are scarcer than the samples

Roughly 2,000 samples collected with a partner hospital produced only about 200 to 250 genuine sepsis cases, because the hospital itself only knows three or four days later. Accuracy sits at 90 to 95 per cent depending on the sample with false negatives under 3 per cent, and he wants 2,000 to 3,000 true sepsis samples before going to market. The instructive case was an apparent false positive: the model called sepsis where the doctors had not, and the next day the clinician confirmed the patient was in kidney failure. Organ failure is treated as sepsis, so the prediction was right even though the treatment differed.

▶ 28:45 ▶ 29:00 ▶ 30:03

The hardest transition for a professor is the reporting line

For decades he told people what to do and then listened; now he sits across from an airport or an investor and defends himself, so he has turned his own supervision habits inward. Every Friday the team writes down what it did, and at three weeks he audits the month, including his own. The sharpest moment came when a promising side-finding surfaced and he shut his own researcher down: this is not research, it is somebody else's money, and it has to convert into a product. Meanwhile he watches go-to-market talks and marketing lectures online, not to become an expert but to understand the language when the MBAs speak.

▶ 13:52 ▶ 15:20 ▶ 16:21

Indian deep tech is short of patience before it is short of science

He made something close to forty presentations over a year and a half in 2021-22 before anyone was convinced, and concluded the problem was that he was explaining pathbreaking science to people who needed to hear how it becomes money. His broader charge is cultural: a middle-class instinct that protects rather than risks, and investors unwilling to underwrite the five-to-ten-year incubation deep tech gets in the West. The line that lands hardest is his own indictment: 75 years, several Nobel laureates from this country, and not a single technology India can sell to the world.

▶ 9:18 ▶ 9:51 ▶ 11:27

Sequence the market: contraband now, hospitals and the world later

Contraband detection is declared ready and sepsis pushed six to eight months back so the existing position can be consolidated. Against colleagues and defence contacts pushing him to chase international security buyers immediately, he wants six months to a year of Indian field use, a few hundred units and twenty machines watched closely, to learn how the instrument fails in the hands of real users before reputation is on the line abroad. The certifications are already done, and the last three months went into a Hindi interface so security forces can be trained in the language they actually use, with Kannada next for state police in small towns.

▶ 30:48 ▶ 32:42 ▶ 33:54

The numbers, drawn

What the episode measures

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

How much of the scattering you actually collect

degrees of scattered light

As stated in conversation: scattering is a 360-degree phenomenon but is traditionally measured at 90 or 180 degrees, wasting 250-280 degrees of light; the patented approach collects the full circle for higher signal-to-noise.▶ 22:39

The sepsis dataset is mostly not sepsis

samples

As stated in conversation: 'about 2,000' samples of which 'only about 200 to 250' were real sepsis (upper bound shown), against a stated pre-market target of 2,000-3,000 sepsis samples (upper bound shown).▶ 28:45

Matched, or suspicious

what happens to a molecule nobody has catalogued yet

Umapathy's account of why library matching fails on anything freshly cooked. His example on air is a new fentanyl family coming back as suspicious and close to fentanyl, which leaves the officer a call to make. Without that, he says, they would be the same as the rest of the world.▶ 21:05

From the prize dinner to the baggage belt

the dated stops on the way out of the lab

The stops Umapathy dates himself. The trial produced the two finds he keeps returning to: a painter's unlabelled white bottle that read as flammable until the man admitted it was thinner, and a suspicious balm that resolved to camphor, both in under five seconds.▶ 2:58

Worth keeping

Lines that stay

One of the things I feel bad about is that in 75 years we don't have a single technology we can sell to the world.

▶ 9:51

We are not doing research here — don't forget, it's somebody else's money, and we need to make sure we convert it into a product.

▶ 15:20

Even if you make a new fentanyl family, it will come and tell you: suspicious, close to a fentanyl family. Then you know you have serious suspicion.

▶ 21:05

There's nothing like developing a technology with the users telling you what they want.

▶ 28:32

Maybe 10 years, 20 years, 30 years down the line, I want the world to recognise Gyrophint as an Indian product, an Indian idea, an Indian technology.

▶ 36:21

Clips that travel

Short on time? Start here

Academics weighing whether to spin a lab out

From an abandoned suitcase to a patent

The Delhi committee question that started it, the Monte Carlo rewrite that made it plausible, DRDO instruments, and the brother who took over the books.

4:14 → 8:15 · 4 min ▶ Watch clip

Deep-tech founders raising a first institutional round

Forty pitches and India's patience problem

A scientist admitting he could not convince a single investor, the coaching that fixed it, and the 75-year indictment that follows.

8:15 → 11:59 · 4 min ▶ Watch clip

Engineers evaluating detection and sensing claims

Suspicious, not matched, and the 360-degree patent

The technical core in four minutes: property-trained AI versus library matching, then the two patents that make the signal good enough to trust.

20:16 → 23:54 · 4 min ▶ Watch clip

Aviation and border-security buyers evaluating screening tools

Three months at Bangalore airport

Thinner in a painter's bottle, camphor in a balm, a five-second scan for land borders, and an airport that wrote to the ministry and lent its screeners as designers.

23:54 → 28:45 · 5 min ▶ Watch clip

Diagnostics teams building models on small clinical datasets

When the model was right and the doctor was not

The sepsis numbers, the false-negative rate, and the kidney-failure case that turned an apparent error into validation.

28:45 → 30:48 · 2 min ▶ Watch clip

Glossary

The jargon, unpacked

Raman spectroscopy

Reading a material by the tiny frequency shift it imprints on scattered laser light. Unlike X-ray or ultrasound, the shift identifies what a substance is chemically, and it can be read through a box or bottle.

Rayleigh scattering

The far stronger, unshifted scattered light that swamps a Raman measurement, along with background noise and stray sources; separating it out mathematically is one of the group's long-standing specialities.

Hyperspectral imaging

Treating a spectrum as a picture, frequency and amplitude across many channels rather than three colours, which is what let the team feed scattering signals to image-style AI models.

Monte Carlo simulation

Statistical modelling of how photons travel and scatter through a medium; the standard formulation assumes the frequency comes back unchanged, so it had to be rewritten for the frequency-shifted Raman case.

Response function

The behaviour a material's physical and chemical properties produce: for an explosive, the blast or the burn. Training on that behaviour rather than on a fingerprint is what lets the system flag compounds it has never seen.

Sepsis

A life-threatening immune reaction to infection, diagnosed slowly. The partner hospital typically confirms only three to four days later, which is why fast detection from body fluids is worth chasing.

False negative

A real case the model calls clean, the error that matters most in both screening and diagnosis; Gyrophint puts its sepsis false-negative rate at under 3 per cent.

DRDO

India's Defence Research and Development Organisation, whose laboratories in Pune and Bangalore took the two pre-startup instruments built from this research.

Connections

If this resonated, go here next

EP 25 · Deep tech

Full transcript

The whole conversation, searchable

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