Capri Global Capital Data Scientist Interview: Questions & Prep (2026)

Capri Global Capital is a listed NBFC with a focus on SME loans, affordable housing finance, and construction lending. Their data science team is embedded in credit, risk, and collections functions, building models that directly affect real lending decisions. As of July 2026, the company has 17 open Data Scientist roles, a sign of active hiring momentum.

Candidates typically report a 3-4 stage process: a recruiter screening call, a technical round (often a take-home case or live coding problem), a domain-focused interview on credit and risk modeling, and a final round with a senior leader or business stakeholder. The technical bar is applied and business-grounded. Expect fewer pure algorithmic puzzles and more questions that test your understanding of how an NBFC makes money and manages risk. If you have prior BFSI (banking, financial services, insurance) experience, lead with it at every opportunity.

These questions reflect the role profile and domain patterns commonly reported by candidates at Capri Global Capital. Prepare honest, specific answers for each.

1. How would you build a credit risk scorecard from scratch for SME borrowers at Capri Global?
2. Loan default datasets are heavily imbalanced. How do you handle class imbalance in your modeling approach?
3. Walk us through a machine learning model you took from idea to production.
4. How would you segment Capri Global's borrower base for targeted cross-sell or upsell campaigns?
5. Explain precision and recall. In the context of predicting loan defaults, which metric would you prioritize and why?
6. What feature engineering steps do you apply when working with financial bureau data like CIBIL or Experian?
7. How would you detect fraudulent loan applications at the application stage, before disbursement?
8. A model you deployed is now showing a spike in prediction errors. How do you investigate and fix it?
9. How do you explain a complex ensemble model's output to a collections or risk team that has no data background?
10. What is survival analysis, and how would it help in modeling borrower repayment behavior over time?
11. How do you ensure a credit model is not unfairly discriminating against a particular borrower segment?
12. You have very limited labeled data for a new loan product. How do you still build something useful?

Q: Walk us through a machine learning model you took from idea to production.

*Situation:* At my previous company, the collections team had no systematic way to prioritize which delinquent borrowers to call first. Agents worked through lists in no particular order, spending equal time on accounts unlikely to pay and those with a strong chance of recovery.

*Task:* I was asked to build a model that scores each delinquent account by its probability of recovery within a defined window, so the team could prioritize their daily call list.

*Action:* I pulled historical collections data, cleaned it, and engineered features like days past due, payment history pattern, loan-to-income ratio, and contact channel preference. I trained a gradient boosting classifier, tuned it with cross-validation, and worked with the engineering team to wrap it into a daily batch scoring job.

*Result:* The collections team used the score to prioritize their call list. Recovery rates improved within the first month of rollout, and agents reported spending less time on accounts unlikely to pay in the short term.

---

Q: A model you deployed is now showing a spike in prediction errors. How do you investigate and fix it?

*Situation:* Several months after deploying a fraud detection model at a previous role, our false positive rate climbed sharply, flagging legitimate customers as fraudulent and creating friction in the application process.

*Task:* I owned the model monitoring dashboard and was responsible for diagnosing and resolving the issue quickly.

*Action:* I first compared the current input feature distributions against the training data distributions. I found that a bureau data vendor had quietly changed how they reported one key variable. I also checked for shifts in application volume and borrower mix. I retrained the model on more recent data with the corrected feature encoding, ran shadow scoring for about a week to validate, then deployed with updated decision thresholds.

*Result:* The false positive rate returned to normal within a few weeks of the fix going live. I also set up automated distribution monitoring alerts so future vendor-side changes would surface faster.

---

Q: How do you explain a complex model's output to a non-technical stakeholder?

*Situation:* Our credit risk model used a stacked ensemble, and the underwriting head wanted to understand why the model was rejecting a particular borrower segment at a higher rate than expected.

*Task:* I had to communicate the reason clearly enough that the business could decide whether to adjust underwriting policy.

*Action:* I used SHAP values to identify the top three features driving rejections in that segment. I then translated those features into plain business language. Instead of saying 'feature X has a SHAP value of 0.3,' I said 'borrowers in this segment most often get rejected because their monthly cash flow shows irregular deposits over the past several months, which the model treats as income instability.' I summarized everything in a one-page visual with no jargon.

*Result:* The underwriting head could clearly see the model logic and decided to collect additional bank statement data for those applicants to reduce ambiguity, rather than loosening the rejection threshold.

For 'how would you approach X' questions, use this five-step structure. It maps directly to how Capri Global's risk and credit teams think:

1. Clarify the business objective: what decision does this model support?
2. State what data you would need and where it comes from
3. Outline your modeling approach and why it fits the problem
4. Describe how you would evaluate success (go beyond accuracy: AUC-ROC, KS statistic, Gini, precision-recall curves)
5. Cover deployment, monitoring, and retraining plan

For behavioral questions, use the STAR method: Situation, Task, Action, Result. Keep it specific. Avoid vague outcomes like 'the team improved.' Use directional results ('recovery rate went up,' 'false positives dropped') even if you cannot share exact figures.

For domain questions on credit risk, anchor your answer in CRISP-DM: Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, Deployment. This structure signals maturity to NBFC interviewers who are used to structured data science processes.

For explainability questions, mention SHAP or LIME by name, and show you can translate technical output into a business narrative. Capri Global's data scientists work closely with non-technical stakeholders, so this skill is evaluated seriously.

Credit and risk domain knowledge. Interviewers at NBFCs want to know you understand how a loan book works, what bureau data contains (CIBIL score, DPD, NPA classifications), and what drives default. If you have BFSI experience, open every relevant answer with it. If you do not, show you have studied the domain independently before the interview.

Applied ML over theoretical depth. Capri Global is building models for real lending decisions. They want evidence that you have shipped models, not just trained them in a Jupyter notebook. Walk through what broke, what you fixed, and what happened after deployment.

Clear communication with business teams. Data Scientists at Capri work closely with underwriting, collections, and product functions. Interviewers often test this directly by asking you to explain a model to a non-technical person. Practice translating ML concepts into business language before the interview.

Ownership beyond model training. Candidates who describe monitoring pipelines, drift detection, and retraining cycles stand out over those who stop at training accuracy. Show that you think of the model as a live product, not a finished artifact.

Intellectual honesty. If you have not worked on a specific problem type, say so clearly and explain how you would approach it. NBFCs deal with regulated, high-stakes lending decisions. Interviewers value candidates who know the limits of their own knowledge.

Week 1: Build domain foundation. Study credit risk fundamentals: how scorecards are built, what bureau variables like CIBIL score and DPD mean, and how NBFCs price and manage default risk. Read Capri Global Capital's annual report and recent investor presentations to understand their loan book, NPA levels, and strategic priorities. This gives you genuine context when domain questions come up.

Week 2: Refresh applied ML skills. Practice building classification models (logistic regression, gradient boosting, random forest) on imbalanced datasets. Review SMOTE, class weighting, and threshold tuning. Prepare walkthroughs of two or three end-to-end projects you have actually completed, including what went wrong and how you fixed it. Practice SQL window functions, aggregations, and joins on transaction-level data, since these are commonly tested.

Week 3: Practice communication and behavioral answers. Prepare STAR answers for questions on model failure, stakeholder communication, and ambiguous problem framing. Do at least one mock technical interview with a peer or mentor. Practice explaining SHAP values and model outputs in plain language out loud, not just in your head.

The day before: Review survival analysis basics, any SQL edge cases you are shaky on, and your own resume projects. Candidates typically report that Capri Global interviews run fairly long, so pace yourself and rest well.

1. Skipping domain context. Saying 'I would train an XGBoost model' without explaining why that fits a credit risk problem signals shallow thinking. Always connect your technical choice to the business context.

1. Over-indexing on accuracy. Accuracy is a misleading metric for imbalanced loan datasets. If you do not mention AUC-ROC, KS statistic, Gini, or precision-recall curves unprompted, interviewers will notice.

1. Stopping at model deployment. Many candidates describe building and shipping a model but say nothing about what happens after. Capri Global wants people who own the model in production: monitoring, drift detection, and retraining.

1. Vague STAR answers. Saying 'the team improved performance' without a concrete outcome weakens your answers. Use honest, directional results: 'recovery rate improved in the first month,' 'false positive rate dropped after retraining.'

1. Not asking questions at the end. Candidates who ask nothing signal low curiosity. Ask about the data infrastructure, what the team is currently building, or how model outputs feed into underwriting policy. These questions also show domain awareness.

1. Treating an NBFC interview like a tech company interview. Capri Global's interviewers care about regulated, high-stakes lending decisions. Answers that are purely ML-technical without business grounding will underperform.