Machine Learning Engineer Interview Questions India 2026: ML Fundamentals, Coding, and System Design

Machine learning engineer interview topics:

1. ML fundamentals: supervised learning (regression, classification), unsupervised learning (clustering, dimensionality reduction), bias-variance trade-off, overfitting vs underfitting

1. Model evaluation: classification metrics (accuracy, precision, recall, F1, AUC-ROC), regression metrics (MAE, RMSE, R-squared), cross-validation

1. Algorithms: linear regression, logistic regression, decision tree, random forest, XGBoost/gradient boosting, k-means, K-NN, SVM

1. Deep learning: neural network architecture (layers, neurons, activation functions), backpropagation, optimisers (SGD, Adam), CNNs, attention mechanism

1. MLOps: feature store, training pipelines, model versioning, A/B testing, model monitoring (data drift, concept drift)

1. Coding: Python (numpy, pandas, scikit-learn), SQL for feature engineering, LeetCode medium problems (arrays, hashmaps, graphs)

1. System design: recommendation systems, fraud detection, real-time ML serving, feature stores

Algorithm questions:

1. 'Explain gradient boosting in plain English': Boosting builds models sequentially; each new model focuses on the examples the previous models got wrong (by fitting to the residuals); combining many weak models (shallow decision trees) creates a strong ensemble. XGBoost and LightGBM are the most popular gradient boosting implementations for tabular data in India.

2. 'When would you choose random forest over XGBoost?' Random forest: less hyperparameter tuning, faster training, robust to outliers, interpretable feature importances, good baseline. XGBoost: typically higher accuracy on tabular data, handles missing values natively, has regularisation to prevent overfitting. Choose XGBoost when maximising accuracy is the priority; choose random forest when you need a quick, reliable baseline.

3. 'Explain the difference between L1 and L2 regularisation': L1 (Lasso): adds the sum of absolute values of coefficients to the loss; can set coefficients to exactly zero (automatic feature selection). L2 (Ridge): adds the sum of squared coefficients; shrinks all coefficients but rarely to zero. Prefer L2 when all features are relevant; prefer L1 when you want a sparse model.

4. 'What is a confusion matrix and how do you use it to choose a model?' TP/FP/TN/FN matrix. For fraud detection: minimise false negatives (missed fraud) even at the cost of more false positives (legitimate transactions flagged). For spam filtering: minimise false positives (legitimate mail in spam) even at the cost of some missed spam. Choose the metric (precision, recall, F1) based on which error type is more costly.

Model evaluation interview questions:

1. 'When would you use AUC-ROC vs precision-recall curve?' AUC-ROC: area under the TPR vs FPR curve. Useful when classes are roughly balanced and you want to measure the overall discriminative ability of the model across all thresholds. Precision-recall curve: more useful when the positive class is rare (high class imbalance). Fraud detection with 0.1% fraud rate: AUC-ROC can look impressive (0.95+) even if the model never correctly predicts fraud; the precision-recall curve would expose this.

2. 'What is cross-validation and why do you need it?' Cross-validation: split the training data into K folds; train on K-1 folds and validate on the remaining fold; repeat K times; average the validation scores. Purpose: gives a more reliable estimate of generalisation error than a single train-test split (reduces variance of the evaluation estimate). Use K=5 or K=10 for standard tabular ML.

3. 'How do you handle class imbalance?' Options: oversampling the minority class (SMOTE: Synthetic Minority Oversampling TEchnique), undersampling the majority class, cost-sensitive learning (set class_weight='balanced' in scikit-learn, which increases the penalty for misclassifying the minority class), collecting more data for the minority class, or changing the evaluation metric (use F1/recall instead of accuracy).

4. 'How do you select features for a model?' Filter methods: correlation with the target (Pearson or Spearman), mutual information. Wrapper methods: recursive feature elimination (RFE) with cross-validation. Embedded methods: L1 regularisation (Lasso), feature importances from tree models (Random Forest, XGBoost). Domain knowledge: always complement statistical methods with subject-matter understanding of which features should matter.

ML system design questions:

1. 'Design a recommendation system for Flipkart': Two stages: candidate generation (retrieve top 1000 candidates from millions of items using collaborative filtering or a two-tower neural network), ranking (score the 1000 candidates with a more complex model using user-item interaction features, price, inventory, margin). Offline: train models on historical interaction data. Online: serve candidates and ranking in under 100ms using pre-computed embeddings and a real-time feature store.

2. 'How would you detect fraudulent transactions at Razorpay in real time?' Streaming pipeline: transactions published to Kafka; ML model scores each transaction within 200ms; features: amount deviation from user history, merchant category risk score, device fingerprint anomaly, time-of-day pattern; model: GBDT trained on labelled fraud data with class imbalance handling (SMOTE or cost-sensitive learning); decision: score above threshold -> flag for review or block; post-review: feed confirmed labels back to retrain.

3. 'How do you handle model drift in production?' Detect: monitor input feature distributions (data drift) and output prediction distributions (concept drift) using Evidently or custom statistical tests (KS test, PSI). Respond: automated retrain trigger when drift exceeds threshold, champion-challenger testing for new model, shadow mode deployment before full cutover.

MLE salary in India (2024-25):

Junior MLE (0-2 years): 10-20 LPA Skills: Python (numpy, pandas, scikit-learn), SQL, ML fundamentals, basic MLOps (model serving with Flask/FastAPI)

Mid-level MLE (2-5 years): 20-40 LPA Skills: deep learning (PyTorch or TensorFlow), distributed training, feature stores (Feast, Tecton), model monitoring, Kubernetes for ML workloads

Senior MLE (5-8 years): 40-70 LPA at top Indian product companies Skills: end-to-end ML platform design, large-scale model training (multi-GPU, distributed), recommendation system or NLP/CV specialisation

Staff MLE (8+ years, FAANG India): 70+ LPA TC

Top MLE employers in India: Flipkart (recommendations, pricing, fraud), Swiggy (delivery time prediction, restaurant ranking), Razorpay (fraud detection, credit scoring), Meesho (product categorisation), CRED (credit risk), Juspay (payment routing), and FAANG India AI teams (Google, Amazon, Microsoft).