Machine Learning Engineer Interview Questions India 2026: Complete MLE Guide

Five question types in Indian MLE interviews: (1) ML theory and algorithms ('Explain gradient boosting: what is it, how does it differ from random forest, when would you use each?', 'What is regularisation? Explain L1 vs L2 penalty and their effect on model sparsity.', 'How does XGBoost handle missing values differently from vanilla gradient boosting?'). (2) Applied ML and feature engineering ('You are building a churn prediction model for Swiggy Pro subscribers. What features would you engineer and why?', 'How would you handle a training dataset where the positive class is 0.1% of the data?'). (3) Coding and data manipulation ('Write a Python function to compute the rolling 7-day average of daily orders, handling gaps in the time series.', 'Using pandas, find all users who made their first purchase in January but did not purchase in February.'). (4) ML system design ('Design a fraud detection system for PhonePe UPI transactions that scores each transaction in under 50ms.', 'Design a product recommendation system for Flipkart's homepage that personalises for 100 million users.'). (5) MLOps and deployment ('How would you detect if a fraud model's performance is degrading in production?', 'Explain what training-serving skew is and how a feature store prevents it.').

ML system design framework (the 6-step approach): (1) Problem framing: define the ML problem precisely (classification, regression, ranking, clustering), identify the business metric (fraud: precision-recall tradeoff; recommendation: CTR, conversion rate, average order value), and clarify constraints (latency SLA: 50ms for fraud, 200ms for recommendation; freshness requirement: how quickly must the model incorporate new data?). (2) Data: what data is available, what data needs to be collected, how is it labelled, and how does the data distribution evolve over time (concept drift)? (3) Feature engineering: which features are most predictive, how are they computed (real-time vs batch), and how are they stored and served consistently (feature store). (4) Model selection: gradient boosting (XGBoost, LightGBM) for tabular data; neural networks for embedding learning from high-cardinality IDs; two-tower models for large-scale retrieval. (5) Serving architecture: real-time vs batch scoring; model server (TorchServe, Triton Inference Server, or custom FastAPI); caching pre-computed scores for high-traffic items; A/B testing infrastructure. (6) Monitoring: feature distribution drift (KS test, PSI), prediction score distribution shift, model performance on a labeled holdout set updated weekly, and feedback loop design.

MLOps technical depth expected at senior MLE level: Feature stores (Feast, Tecton, or custom Redis/Cassandra implementation: the feature store computes and caches features so that training and serving use exactly the same feature values, preventing training-serving skew, one of the most common causes of model underperformance in production). Model serving (real-time: REST API with FastAPI or Flask behind a Kubernetes deployment; batch: Spark or Ray for scoring millions of records hourly; model versioning: blue-green deployment for switching between model versions without downtime; shadow scoring: run new model in parallel with production model before cutover). Model monitoring (distribution drift detection: Kolmogorov-Smirnov test for continuous features, chi-squared test for categorical features, Population Stability Index (PSI) for scorecard models; performance monitoring: track AUC, Precision@K, or business metrics on a weekly labelled holdout set). Retraining pipelines (Airflow DAGs or Kubeflow Pipelines for scheduled retraining: trigger on schedule, on drift signal, or on label availability; champion-challenger: route 10% of traffic to the challenger model and compare performance before full rollout). Experiment tracking (MLflow or Weights and Biases: log every training run's parameters, metrics, and model artefacts; model registry for promoting experiments to staging and production).

ML engineer salary ranges: junior MLE (0-2 years): Rs 12-30 LPA. Mid-level (2-5 years): Rs 30-70 LPA. Senior MLE (5-8 years): Rs 65-130 LPA. Principal MLE / ML Architect (8+ years): Rs 120-220 LPA. Top MLE employers in India: Google India (Rs 60-180 LPA mid-senior: search quality, advertising relevance, YouTube recommendations). Amazon India (Rs 45-120 LPA: Alexa India NLP, AWS SageMaker platform, Prime Video personalisation, India e-commerce demand forecasting). Flipkart (Rs 35-80 LPA: one of India's most mature ML environments for recommendation, search ranking, and supply chain optimisation). Swiggy (Rs 30-70 LPA: ETA prediction, demand forecasting, driver supply optimisation, food recommendation). PhonePe (Rs 35-80 LPA: real-time fraud detection, credit risk scoring for lending products). The MLE skill set is the most in-demand combination in Indian tech: software engineering plus ML knowledge opens opportunities at every company with data-driven products.