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

Episode 68 · Fintech · 52 min

Graded on a curve built for salaried India

India's credit bureaus grade everyone on one scale, so a mason paid four days late loses a hundred points and lands in the 450–650 bin with everyone else. Kaleidofin's answer is a supervised machine-learning score trained on 1.5 crore customer records and stress-tested against demonetisation, floods and droughts — 16 lakh customers underwritten, over ₹6,500 crore disbursed, collections above 99%.

Credit & lending Data & digitisation AI & machine learning

Puneet Gupta

Co-founder, Kaleidofin · with Vishal Krishna

52:19

Said in this episode

▶ 14:52

₹6,500 cr

Disbursed on ki score reports

Partners have given out around 1.2 million loans using the model; the book Vishal cites at ₹3,000 crore during COVID had roughly doubled in about eighteen months.

▶ 15:10

99%+

Collection rate through the pandemic

Collections on loans underwritten by the model stayed above 99% across the disruptions of both COVID waves — Gupta's evidence that it predicts default well.

▶ 12:34

1.5 crore

Customer records behind the model

Seven to ten years of credit-bureau trade lines plus partners' own loan data for over 1.5 crore customers went into training the supervised score.

▶ 7:19

12 of 135 cr

Indians who have ever paid digitally

A statistic Gupta attributes to a Nandan Nilekani presentation — roughly one in six or seven of the country's 70–90 crore adults has ever made one digital financial transaction.

▶ 11:48

450–650

Where unorganised-sector scores bunch

Most of these customers sit between 500 and 650, sometimes as low as 450, while banks treat anything below 700 as a bad score and below 750 as not a great one.

▶ 44:39

97%

Women among Kaleidofin's customers

97% of all customers and about 94% within the nano-entrepreneur segment are women, though Gupta notes the spouse tends to join the business once it scales.

The brief

The argument in sixty seconds

Gupta's claim is that the credit bureau is a relative-grading machine built for salaried borrowers: put a mason, a driver and a management trainee on one curve and the mason's ₹1,000 instalment, three days late because his employer had no cash that week, costs him a hundred points. Whole cohorts bunch between 450 and 650 while banks refuse to look below 700, so the score stops discriminating at exactly the point it should. Kaleidofin's ki score answers with segmentation rather than sympathy — a tree of models that first asks whether you have any credit file at all (bureaus simply return 'no score found', Kaleidofin maps you to a look-alike persona), then differentiates by occupation, by state, by how long you waited between loans, and finally by amount: not creditworthy in the abstract, but creditworthy for ₹10,000. It was trained on seven to ten years of data for over 1.5 crore customers and modelled against real shocks — demonetisation, floods, droughts — and partners have since disbursed over ₹6,500 crore on its reports with collections above 99% through both pandemic waves. The second claim is stranger: savings and insurance are credit products, because an insured asset and a small monthly buffer let a household absorb the bad month without missing an instalment. Around it sit the constraints — roughly 12 crore of 135 crore Indians have ever made a digital payment, a paper mandate still carries the bulk of collections, an NBFC licence application sits with the RBI, and the team argues internally over whether a pincode is a signal or a redline.

Worth your time if you are

Lenders underwriting borrowers with no bureau file

Data scientists building models on thin data

Impact investors weighing returns against reach

Founders whose customers do not own smartphones

Policy people arguing about KYC and the BC model

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.

Credit & lendingData & digitisationAI & machine learningPayments & fintechImpact & outcomesRegulation & policy

01Cold open: accounts opened, accounts unused 0:00 Vishal frames financial inclusion with the government's own numbers — 192.1 million zero-balance Pradhan Mantri accounts, 165.1 million debit cards, a ₹30,000 life cover — then asks the question those numbers dodge: how do people actually use the accounts, and who pays the money back? 02ICICI, IFMR, and the phone as branch 2:28 Gupta joins ICICI Bank in 2001 in a grant-making department because nobody yet believes finance for the poor can be commercial, builds a ₹5,000 crore microfinance book in five years, moves to Chennai to build IFMR Trust's branch network — now branded Dvara and Northern Arc — and by 2017 concludes that Jan Dhan, Aadhaar and cheap smartphones have made the branch itself obsolete. 03Voice, paper and the one-time mandate 5:48 Kaleidofin Pay exists because only around 12 crore of 135 crore Indians have ever made a digital financial transaction; it combines voice, paper mandates and digital rails so that a feature-phone customer with a working bank account can be direct-debited instead of visited for cash. 04Why bureau scores bunch the poor together 10:34 A ₹1,000 instalment three days late costs a hundred points, and a mason whose employer delays his wages looks identical to a defaulter — so unorganised-sector customers pile up between 450 and 650 while banks refuse to look below 700, and the score stops being a differentiator. 05Trained on demonetisation, floods and droughts 12:20 ki score is a supervised machine-learning model built on seven to ten years of data for over 1.5 crore customers, ranking households by how they behaved through real shocks rather than by a single grade; it has underwritten 16 lakh customers and around 1.2 million loans, over ₹6,500 crore disbursed, with collections above 99%. 06Equity raises are charging points 15:53 With awards, 230 districts and 14 states behind them, Gupta likens fundraising to hunting for a charger every hundred kilometres — twelve months rather than twelve hours — argues a social enterprise that cannot show unit economics will only ever attract one kind of capital, and puts the new round into a credit business and a zero-cost co-branded bank account built with a licensed bank. 07What a nano entrepreneur actually is 19:27 Under ₹1 crore of annual turnover, roughly ₹3 lakh of net income a year, fewer than five employees, no GST registration and receipts mostly in cash — about 95% of India's MSME cohort, and invisible to anyone who underwrites on documents. 08The moat is the customers nobody wants 22:10 Rivals chase borrowers who already have bureau files, documents and smartphones — some mining SMS inboxes for alternate data — while a large share of Kaleidofin's customers own no smartphone at all, which is exactly why their behaviour sits in no competitor's dataset. 09Savings and insurance make better borrowers 23:48 Business risk is built into the word entrepreneur, so pure finance says never fund one — Gupta's answer is to sell the household savings, insurance and investments alongside credit, so an insured asset and a small monthly buffer turn a vicious borrowing cycle into a virtuous one. 10B2B2C, with an NBFC licence pending 26:21 Partners own field distribution and collections while Kaleidofin supplies underwriting IP and technology; the book cited at ₹3,000 crore during COVID is now ₹6,500 crore, and an application for a non-bank finance licence sits with the RBI while banks and NBFCs fund the loans in the interim. 11A tree of models, not one number 28:37 Instead of a single relative grade the score branches — a look-alike persona for customers the bureau returns 'no score found' for, then occupation, then state — and ends by underwriting an amount: not a good credit risk for ₹1 lakh, but a very good one for ₹10,000. 12Shop photos, GST invoices, account aggregators 32:27 Loan applications, pictures of the shop, GST invoices and bank statements give the model more per-customer information than a bureau holds, and Gupta argues India is now the best place in the world to build this — account aggregators, UPI, ONDC and consent rules that leave data ownership with the customer. 13Statisticians, credit people and activists 36:47 The hiring problem is data scientists who can drive Python but cannot say why a technique was chosen, so the team deliberately mixes people who understand statistics, people who understand credit, and social-equality activists who argued age, gender, caste, religion and pincode out of the model. 14Stability is the strength and the cost 39:24 A strong regulator is why investors trust the Indian system, Gupta says, but the same rigidity leaves almost no room to experiment: a non-bank reaches the customer only as a business correspondent, KYC has no portability, and his ask is licensing relief until an innovation proves itself at scale. 1597% women, and beautiful Mondays 44:12 Women are 97% of all customers and about 94% in the nano-entrepreneur segment, five of eight board seats and the CTO's chair; Gupta closes on investors whose board meetings discuss nothing but the customer, on the guilt that started him, and on the Steve Jobs biography his son told him to finish.

Takeaways

Ideas to carry out of this hour

The bureau grades the poor on a curve built for the salaried

A credit bureau puts you, me, our drivers and our maids on one relative scale, and relative grading punishes the bottom of the class: once a very bright student takes the A+, everyone else slides toward a C. For an unorganised-sector customer the hiccups are rarely defaults at all — a payment that missed its scheduled date because a contractor paid four days late — yet a ₹1,000 instalment three days late can knock fifty, a hundred, a hundred and fifty points off. The result is a cohort bunched between 450 and 650 against banks that call anything under 700 a bad score, which means the number no longer separates anyone from anyone.

▶ 11:03 ▶ 11:34 ▶ 29:37

Underwrite the amount, not the person

A bureau score of 700 or 800 tells a lender nothing about size: it does not say whether this customer can repay ₹10,000 or ₹1 lakh. ki score is built as a tree instead — customers with no file go to a look-alike model that maps them onto a persona whose risk is known, rather than being returned as 'no score found'; customers with history are then differentiated from the average persona by occupation (mason, plumber, farmer) and by geography (Tamil Nadu against Bihar). The output is deliberately amount-specific, and Gupta argues that is not empathy but the segmentation that should have been done in the first place.

▶ 30:22 ▶ 31:06 ▶ 32:11

Train the model on shocks, not on good years

Kaleidofin modelled default events rather than steady-state repayment: who defaulted through demonetisation, who paid first, who came back three or four months later once the physical cash crunch eased. Floods and droughts do the same work for farmers — and Gupta concedes the rank order may be partly luck, since one farmer's pincode simply did not flood, but a farmer whose farm floods is genuinely more vulnerable. The stated design goal follows from that: a model that is explainable, robust, and does not only work when markets are good.

▶ 13:36 ▶ 14:06 ▶ 39:06

Savings and insurance are credit products

Business risk is built into the word entrepreneur, so on pure finance logic a lender should never fund one — except that this customer has no other option at all. Gupta's answer is to underwrite the household rather than the loan: insure the primary asset so the business is not wiped out with it, and get the entrepreneur to save a small slice of monthly receipts so a lean month can be covered without missing an instalment or skipping meals. The claim is that resilience precedes creditworthiness — a household with an emergency buffer borrows more cheaply, earns more, and moves off the vicious cycle of borrowing at ever higher rates to service the last loan.

▶ 24:01 ▶ 24:50 ▶ 25:51

The customers nobody else wants are the moat

Everyone else is competing for borrowers who already have documented income, a bureau score and a smartphone — including the digital lending apps that read SMS inboxes for alternate data. A very large share of Kaleidofin's customers do not own a smartphone at all, so the behavioural data the company accumulates as it lends to them exists in nobody else's system. Gupta reads the ₹6,500 crore built in eighteen months, against almost no competition and with rivals asking to partner instead, as evidence of how large and how unserved the segment is.

▶ 22:35 ▶ 23:03 ▶ 23:18

Hire the activist alongside the statistician

Gupta's hiring complaint is data scientists who know how to drive Python or a statistical tool but cannot say why a technique was chosen, what the output means, or why a variable should matter — people who place their confidence in the model because the computer said so. Kaleidofin deliberately mixed in people who understood credit and carried a hypothesis about each variable, and social-equality activists who fought hardest to keep age, gender, caste, religion and pincode out of the model, and who dismissed the 'negative pincode' lists banks maintain as meaningless. The question they force on the team is whether the model is digitally redlining the wrong people.

▶ 37:03 ▶ 38:16 ▶ 38:51

Regulatory stability is both the strength and the cost

Gupta credits India's regulators for a system where entities do not go ballistic and confidence rarely collapses — the reason investors are comfortable in the first place. The same rigidity leaves almost no space to experiment: to offer a customer an account a non-bank is at best a business correspondent passing information to the issuer bank, unable to change a single feature; KYC has no portability because the rules require each entity to do it itself, so centralised KYC does not work in practice. His proposal is to absolve an innovator of licensing requirements while a well-capitalised banking partner carries the risk, and to regulate only once the innovation has proven itself at scale.

▶ 40:24 ▶ 41:23 ▶ 43:24

Reach is a payments problem before it is a credit problem

Citing a Nandan Nilekani presentation, Gupta puts the number of Indians who have ever made a single digital financial transaction at roughly 12 crore out of 135 crore people — about one in six or seven adults. Kaleidofin Pay is built for the other five: customers with no smartphone, or with a debit card they have never activated, are enrolled through a one-time mandate and then transacted with by voice, so lenders can direct-debit instead of sending an agent to collect cash. Digital mandates exist and are supported, but the bulk of the volume still runs on paper.

▶ 7:19 ▶ 7:50 ▶ 10:18

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

Credit & lending24%

Data & digitisation16%

AI & machine learning13%

Payments & fintech12%

Impact & outcomes10%

Regulation & policy9%

Computed from the chapter map of this episode.

Where the unorganised sector lands on a bureau score

bureau score

As stated in conversation: these customers typically score 500–650 and sometimes 450–650 (lower bound shown), while most banks call anything under 700 a bad score and under 750 not a great one.▶ 11:48

The book underwritten by ki score

₹ crore disbursed

Vishal cites a ₹3,000 crore figure from earlier reporting on the COVID period; Gupta confirms the book has since moved to about ₹6,500 crore, which he elsewhere describes as eighteen months of building.▶ 26:52

How few Indians have ever paid digitally

crore people

As stated in conversation, attributed to a Nandan Nilekani presentation; the auto-captions drop the unit on the first figure, but Gupta's own gloss — one in six or seven of 70–90 crore adults — makes 12 crore the intended number.▶ 7:19

Worth keeping

Lines that stay

Once you have a super bright person in your management class who gets an A+, everybody else in the class starts to get a C.

— Puneet Gupta ▶ 29:37

From a financial point of view I should never fund an entrepreneur, because they are risky. But this particular customer does not have any other option at all.

— Puneet Gupta ▶ 24:01

My score now says you may not be a good credit risk for a hundred thousand, but you are a very good credit risk for ten thousand.

— Puneet Gupta ▶ 32:11

We had to keep asking ourselves — are we digitally redlining the wrong people?

— Puneet Gupta ▶ 38:51

I actually find it hard to say what is my personal goal and what is my organisation's. As a result, Monday mornings are beautiful.

— Puneet Gupta ▶ 49:20

Clips that travel

Short on time? Start here

Lenders underwriting borrowers with no bureau file

Why the mason loses a hundred points

The relative-grading problem stated plainly, then the shock-event method — demonetisation, floods, droughts — used to rank household vulnerability instead.

10:34 → 14:37 · 4 min ▶ Watch clip

Impact investors weighing returns against reach

What a nano entrepreneur actually is

The segment defined in numbers — turnover, income, headcount — plus why no registration, no GST and a cash economy make it both unreachable and defensible.

19:27 → 23:48 · 4 min ▶ Watch clip

Product leaders bundling credit with everything else

Savings make better borrowers

The episode's least obvious argument: insurance and a small monthly saving are what make a risky entrepreneur bankable.

23:50 → 26:21 · 3 min ▶ Watch clip

Data scientists building models on thin data

A tree, not a single number

The architecture of the score walked through step by step, ending on the shift from creditworthy to creditworthy-for-₹10,000.

28:37 → 32:27 · 4 min ▶ Watch clip

Anyone building models that decide who gets money

Hiring the activist who argues pincodes out

Statisticians versus tool operators, credit people with hypotheses, and the internal fight to keep caste, religion and pincode out of an underwriting model.

36:47 → 39:24 · 3 min ▶ Watch clip

Glossary

The jargon, unpacked

ki score

Kaleidofin's supervised machine-learning credit score for unorganised-sector customers, trained on seven to ten years of data covering more than 1.5 crore borrowers and used to underwrite loans for partner lenders.

Nano entrepreneur

A self-employed business with under ₹1 crore of annual turnover, roughly ₹3 lakh of net annual income and fewer than five employees — around 95% of India's MSME cohort, by Gupta's count.

Look-alike model

The branch of the score used for customers with no credit file: rather than returning 'no score found', it maps the applicant onto a persona whose risk profile is already known.

NACH mandate

A standing instruction that lets a lender debit a borrower's bank account automatically; Kaleidofin supports paper, electronic and UPI versions, and paper still carries the bulk of its volume.

Business correspondent (BC)

The agent model through which a non-bank can bring a customer to a bank; the bank remains the issuer, so the correspondent cannot change anything about the account it distributes.

Account aggregator

India's consent-based framework for sharing bank data between institutions, which Gupta says removes any need to mine a customer's SMS inbox for transaction history.

Digital redlining

Using proxies such as pincode, age, gender, caste or religion to exclude whole groups from credit — the failure mode Kaleidofin's team argued out of its own model.

NBFC

A non-banking financial company. Kaleidofin has applied to the RBI for a licence so it can lend off its own balance sheet instead of only underwriting for partner lenders.

Connections

If this resonated, go here next

EP 64 · Fintech

Full transcript

The whole conversation, searchable

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