KPMG India Machine Learning Engineer Interview: Questions & Prep (2026)

KPMG India currently lists 233 open roles for Machine Learning Engineers (knok jobradar, July 2026). As one of the Big Four consulting firms, KPMG deploys ML across audit analytics, fraud and risk detection, tax compliance automation, and digital transformation programs for enterprise clients. The broader market shows 803 ML Engineer openings across India, with Bangalore leading at 165 openings, followed by Delhi (50) and Hyderabad (27).

KPMG's interview process typically runs across two to four rounds, candidates report, covering a coding or take-home assessment, one or two technical discussions, and an HR conversation. What sets KPMG apart from product-company interviews is the consulting lens: you are expected to frame model choices in business terms, speak to regulatory awareness (especially for banking and insurance clients), and show that you can communicate findings clearly to non-technical stakeholders.

1. Walk us through an end-to-end ML project you built from scratch. What problem did you solve and how did you measure success?

1. How do you handle class imbalance in a fraud detection dataset? Which techniques have you applied and how did you evaluate them?

1. KPMG frequently works with structured, tabular client data. When would you choose gradient boosting over a neural network, and why?

1. How do you explain a model's predictions to a non-technical client or an audit committee?

1. Describe your MLOps experience. How have you managed model versioning, retraining pipelines, and production monitoring?

1. A client's model shows high accuracy but the business outcome is poor. Walk us through how you would diagnose and fix this.

1. How do you ensure data privacy and regulatory compliance (for example, under the DPDP Act) when training models on client data?

1. What is your approach to feature engineering for time-series financial data?

1. How would you design an anomaly detection system for a large bank's transaction data?

1. KPMG serves clients across industries. How do you ramp up on a new domain's data quickly and deliver within a consulting project timeline?

1. Describe a situation where your model or recommendation was rejected by a stakeholder. How did you handle it?

1. How do you keep up with ML research while meeting client delivery deadlines?

Q: Walk us through an end-to-end ML project you built from scratch.

*Situation:* My team was asked to reduce customer churn for a telecom product. Churn was rising but the business had no data-driven way to prioritise retention outreach.

*Task:* I was responsible for building and deploying the churn prediction model end to end, from raw data to a usable output for the retention team.

*Action:* I started with twelve months of usage data, ran exploratory analysis to identify the strongest behavioral signals (drop in data usage, support ticket frequency, billing disputes), and trained an XGBoost classifier. I tuned hyperparameters with stratified cross-validation and evaluated on precision-recall rather than accuracy, since churn events were rare. I then set up a weekly retraining pipeline and built a simple prioritised list output so the retention team could act without touching the model directly.

*Result:* The retention team adopted the output in the first sprint. The business reported a measurable improvement in churn within the first quarter, which was cited internally as the project's main success metric.

---

Q: How do you handle class imbalance in a fraud detection dataset?

*Situation:* I was building a payment fraud classifier where genuine transactions outnumbered fraudulent ones by a wide margin, making accuracy a misleading metric.

*Task:* I needed a model that caught fraud without generating so many false positives that the operations team could not keep up with reviews.

*Action:* I evaluated three approaches side by side: SMOTE oversampling on the minority class, class-weight adjustment in the model, and probability threshold tuning post-training. For each, I measured performance using precision-recall curves and F1 at several operating thresholds. I also involved the fraud ops lead early to agree on the acceptable false-positive rate before finalising the threshold.

*Result:* The class-weight model with calibrated threshold tuning performed best in production. The fraud operations team reported a significant drop in missed cases compared to the legacy rule-based system, and the false-positive rate stayed within the agreed limit.

---

Q: Describe a situation where your model was rejected by a stakeholder.

*Situation:* I built a neural network to score supplier risk for a procurement team. The model performed well in testing but the procurement head called it a 'black box' and refused to use it.

*Task:* I needed to rebuild trust in the model's outputs without replacing the model itself.

*Action:* I added SHAP value explanations to every prediction so the team could see which factors drove each risk score (payment delays, audit findings, geographic concentration). I then ran a two-week parallel test alongside the manual review process and presented side-by-side comparisons showing where the model agreed with and differed from the team's own judgements.

*Result:* After seeing the explanations, the procurement head agreed to a pilot for the top supplier tier. The model was fully adopted within two months, and the team said the transparency was the deciding factor.

For technical ML questions: Start with 'why' (the business or data problem), then 'what' (the method you chose), then 'trade-offs' (what you gave up and why that was acceptable). KPMG interviewers value consulting-style thinking, so always connect your technical choice back to the client's constraints: time to deliver, interpretability requirements, or data availability.

For case-style and domain questions: Use a structured problem-framing approach. Restate the objective, list the data you would need, name two or three candidate methods, explain your selection criteria, and describe how you would validate results before handing over to the client. Candidates report that KPMG interviewers reward structured reasoning over a single 'correct' answer.

For behavioural questions: Use the STAR format (Situation, Task, Action, Result). Keep Situation and Task brief, one or two sentences each. Spend the most time on Action (what you specifically did, not what the team did) and Result (a concrete outcome or a clear lesson). If you do not have an exact metric, describe the qualitative impact and be honest about why you cannot quantify it.

Business fluency: KPMG is a consulting firm first. Interviewers want to see that you can frame an ML problem in terms a CFO, audit partner, or risk officer understands, not just a data science team. If you can only speak in model metrics, you will struggle in this environment.

Production and deployment mindset: Expect questions about monitoring, retraining schedules, and model drift, not just training accuracy. Mentioning specific MLOps tools you have used in real projects signals the kind of maturity KPMG looks for in a consultant-facing role.

Regulatory and data governance awareness: KPMG's clients in banking, insurance, and government require careful data handling. Familiarity with frameworks like the DPDP Act, audit trail requirements, or model risk management guidelines is a genuine differentiator.

Clear communication under pressure: Candidates report that KPMG rounds often include a prompt to 'explain this to a non-technical stakeholder.' Practice translating model outputs into plain recommendations with clear caveats, and avoid jargon even when discussing technical topics.

1. Review core ML concepts over two weeks: gradient boosting, regularisation, evaluation metrics (precision, recall, AUC-ROC), dimensionality reduction, and clustering. Be ready to compare methods and explain your choice for a given scenario.

1. Practice coding on platforms like LeetCode or HackerRank, focusing on data manipulation (pandas, SQL) and medium-difficulty algorithm questions. SQL is frequently tested for data aggregation and transformation tasks.

1. Build your MLOps knowledge: understand model versioning with tools like MLflow, CI/CD pipelines for ML, and monitoring strategies for detecting data drift in production.

1. Develop a consulting mindset: read one or two publicly available KPMG India thought-leadership reports on AI in audit or risk. Understanding how KPMG positions ML to clients helps you frame answers in the language interviewers use.

1. Prepare STAR stories for at least five scenarios: a project you led end to end, a failure or stakeholder rejection, a cross-functional collaboration, a time you handled ambiguous or poor-quality data, and a time you learned a new domain quickly.

1. Practice explaining out loud: record yourself describing a past project or model decision without notes. Review for jargon and check that a non-technical listener could follow along.

For the job search itself, knok checks 150+ job sites nightly, applies to roles matching your resume, and messages HR on your behalf. With 233 open ML Engineer roles at KPMG India alone, having automated coverage helps you stay ahead of the queue.

1. Skipping the 'why' on model choices. Saying 'I used XGBoost' without explaining why it suited the problem signals shallow thinking. Always connect method to context.

1. Ignoring interpretability. KPMG's audit and risk clients often require explainable outputs. Discussing only neural networks without mentioning tools like SHAP or LIME misses a core concern for this firm.

1. No production framing. Describing only Jupyter notebooks or Kaggle projects without any mention of deployment, monitoring, or retraining is a red flag in a consulting ML role where clients expect maintained, documented solutions.

1. Overly academic answers. Citing research papers without connecting them to practical outcomes or client constraints suggests you may struggle under consulting delivery timelines.

1. Not asking questions at the end. Candidates report that failing to ask about the team's ML stack, client industries, or how success is measured on projects is read as a lack of genuine interest. Prepare two or three specific questions before each round.