Kalpa Financial Services

Payments and consumer lending. The unit where the two errors cost different amounts, and where almost every label arrives late or never arrives at all.

Code FIN. Introduced as a transfer domain once a concept has landed on the Retail spine, and as one of the five build-week sub-problems.

Where it sits in the locked ladder

It enters on Week 5 as the first domain transfer, once the propensity model has landed on Retail. The stakeholder is Rohan Desai, head of risk, and his question is the same model in a business where a wrong call costs twenty times more one way than the other. That asymmetry is the whole reason this unit goes first in the rotation.

Build 2, Week 6. Five sub-problems, each with a constraint written into the brief:

#	The sub-problem	The constraint that makes it hard
1	Credit approval	A default costs twenty times a wrongful rejection
2	Fraud detection	Half a percent positives
3	Collections prioritisation	A fixed team, so capacity sets the threshold
4	Churn on the payments app	The target has to be defined before anything is modelled
5	A lending campaign that lifted approvals 9 percent	Cause or mix

The constraint is the teaching. Every one of these is answerable with a model the room already knows, and every one of them is answered wrongly by a leaderboard metric.

---

1. Who runs it, and what each of them wants

Role	Measured on	What they want to hear	What they will resist
Head of Growth	Approved applications and disbursed volume	That approvals can rise	Any model that refuses more people this quarter
Head of Credit Risk	Loss rate on the book	That the book is safe	A loss number computed on a mature cohort rather than a growing one
Head of Fraud Operations	Chargebacks paid, and her analysts' workload	That the alert queue will fit in a day	An alert volume set by a probability threshold rather than by capacity
Head of Collections	Recovery rate by arrears bucket	That the new script works	A comparison that holds debt age constant
Compliance	Findings, and being able to explain a decision	That every decision has a reason a regulator will accept	A reasoning trace nobody has tested

The tension that generates most cards here: growth is measured monthly and risk matures over a year, so the person who wins this quarter is often the reason the next one is bad. Every honest metric in this unit is a metric that makes somebody wait.

---

2. The data estate

What exists	The catch it carries
Transactions, with a chargeback_received flag	The flag lands 45 to 90 days after the transaction, so recent data is labelled clean because nobody has complained yet
Analyst review logs, one disposition per reviewed case	Only cases the model surfaced were ever reviewed, so the labels describe the model's own past taste
Loan applications with bureau attributes at the time of application	Nothing at all for applications that were rejected
Repayment schedules and arrears buckets	Maturity differs by cohort, so a young book has a flattering loss rate by construction
Collections contact logs	The easiest bucket was contacted first, which confounds every before-and-after
The policy document estate	More than one version of some policies is live at once

The absence that defines the unit: rejected applicants have no outcome. Every credit model in the world is trained on people it approved, and evaluated on survivors. Naming that unprompted is a strong interview signal.

---

3. The KPIs, and how each one is gamed

KPI	What people say it is	The denominator that decides it	How it breaks or gets gamed
Approval rate	Approvals divided by applications	Which applications arrived, which is a partner's choice not yours	A partner changing its own pre-filter moves this number with nothing changing on your side
Loss rate	Defaults divided by disbursed	The mature book, cohort by cohort	Grow fast and the denominator outruns the numerator, so losses appear to fall while the book gets worse
Fraud detection rate	Caught divided by total fraud	Total fraud is unknown, since you only see what was reported	Prevention destroys its own labels: a blocked transaction never becomes a chargeback
Alert precision	True fraud among alerts	The top N, where N is what nine analysts can work in a day	A threshold of 0.5 is a number from a textbook. Capacity is the number from the business.
Chargeback rate	Chargebacks divided by transactions	Only settled months	The current quarter always looks like the best one on record
Collections recovery	Recovered divided by outstanding	Within an arrears bucket	Roll a new script out to the freshest arrears first and it will beat the old one every time
Cost of an error	Rarely stated at all	Rupees, on both sides	A blocked good customer leaves quietly and never appears in a fraud report. A missed fraud is visible, so it dominates the conversation and the model.

The KPI this unit exists to teach: value recovered per analyst hour. It forces capacity, precision and cost into one number that a Head of Fraud Operations can actually act on.

---

4. Case families, mapped to modules

Module	Case family here	The teaching version
1 Foundations of AI and Data	Approval and conversion rates across channels with different pre-filters	The denominator lesson with real stakes attached
2 Applied Machine Learning	Fraud detection under extreme imbalance, credit scoring with asymmetric cost, reject inference	Ranking rather than classifying, and a threshold set from a capacity constraint
3 Deep Learning and Neural Networks	Sequence models over a customer's transaction history	Where a feature-engineered baseline should be beaten before anything deeper is justified
4 Natural Language Processing	Merchant name normalisation, dispute narrative classification	Messy short text with high business consequence
5 and 6 Generative AI	Policy and product document retrieval for a servicing team	Where a corpus with two live versions of one policy is the whole lesson
7 Production AI	Drift when underwriting policy changes, and monitoring a model whose labels lag 90 days	Detecting a change you cannot yet measure
8 and 9 Agentic AI	Transaction monitoring and servicing agents, tool authorisation, step-up confirmation	S-FIN-GA-02, which is the hardest card in the bank

---

5. The situations hosted here

Page	Cards
Analytics	S-FIN-AN-10 to S-FIN-AN-12
Machine learning	S-FIN-ML-01 (five hundred alerts a day), S-FIN-ML-04 to S-FIN-ML-06
GenAI and agents	S-FIN-GA-02 (the agent that read the customer's email), S-FIN-GA-08 to S-FIN-GA-10

---

6. What Financial Services must never be used for

1. Any real financial data, ever. The repository is public. Every figure here is generated or invented, and it is labelled as such.
2. A regulatory claim. Cards can say a regulator would ask for an explanation. No card asserts what any actual regulation requires, because nobody in the build has verified it.
3. A fairness case treated as a maths exercise. S-FIN-ML-05 exists because pincode and device reconstruct a protected attribute. That card needs a trainer who can hold the room, and it does not belong on a Week 5 teaching day.