ML Engineer Interview Questions India 2026: Complete Guide

These topics appear in every MLE interview in India regardless of company or seniority level. Bias-Variance Trade-off: high bias means the model is too simple (underfitting: makes strong, wrong assumptions about the data). High variance means the model is too complex (overfitting: memorises training data but fails on new data). The sweet spot is the model complexity that minimises total error = bias^2 + variance + irreducible noise. Practical manifestation: a linear model on a non-linear problem has high bias. A deep neural network trained on 500 examples has high variance. Regularisation: L1 (Lasso) regularisation adds the sum of absolute values of weights to the loss function. It drives some weights to exactly zero, performing feature selection. L2 (Ridge) regularisation adds the sum of squared weights to the loss. It shrinks all weights toward zero but rarely to exactly zero. Elastic Net combines both. Interview question: 'You have 500 features for 1,000 training examples and your model is overfitting. What do you do?' Options: L1 regularisation for feature selection, L2 for weight shrinkage, reduce model complexity, collect more data, or use cross-validation to estimate true generalisation. Class Imbalance: in fraud detection or rare disease diagnosis, positive examples may be 0.1% of the data. Solutions: change the evaluation metric from accuracy to precision-recall AUC or F1. Use class weights in the loss function to penalise false negatives more heavily. Oversample the minority class (SMOTE: Synthetic Minority Oversampling Technique). Undersample the majority class. At the threshold level: adjust the classification threshold below 0.5 to increase recall. Ensemble Methods: Random Forest: train N decision trees on different random subsets of data and features, then aggregate predictions (bagging). Gradient Boosting (XGBoost, LightGBM, CatBoost): train trees sequentially, each correcting the errors of the previous (boosting). Interview question: 'When would you use a Random Forest over XGBoost?' Random Forest is more parallelisable and less prone to overfitting without careful tuning. XGBoost typically achieves higher accuracy on tabular data with proper hyperparameter tuning.

ML system design is the most important topic that separates junior and senior MLE candidates in India. The framework for any ML system design question: Business Problem: what decision is the ML model making and for whom? Data: what data exists, how is it collected, is it labelled, what is the label quality? Features: feature engineering from raw data, feature store, online vs offline features. Model: algorithm selection and justification, training infrastructure, offline evaluation (metrics, cross-validation strategy). Serving: latency requirements, batch vs real-time inference, model versioning. Monitoring: data drift, concept drift, model performance decay, feedback loops. Worked example: 'Design a fraud detection system for UPI transactions at 100 million transactions per day.' Business problem: identify fraudulent transactions in real time before the payment completes. Data: historical transactions with fraud labels (from chargebacks and manual review), approximately 0.05% positive rate. Features: transaction amount, merchant category, time of day, device fingerprint (new device for this user?), velocity features (number of transactions in last 10 minutes for this user), geographic anomaly (payment in a city the user has never transacted from). Model: XGBoost for the main model (handles tabular features well, interpretable, fast inference). Ensemble with a rules engine for high-confidence fraud patterns (a transaction of Rs 2 lakh to a new merchant from a new device at 3 AM). Serving: online inference at <100ms latency (the payment API calls the fraud model synchronously). Feature store: pre-compute velocity features for each user and update in Redis in real time as transactions arrive. Monitoring: monitor false positive rate (legitimate users blocked), false negative rate (fraud that went through), and the distribution of model inputs for data drift.

The rise of large language models has significantly changed MLE interviews in 2026. Companies building AI-first products (Sarvam AI, Krutrim, Ola AI) require deep LLM engineering skills. Even companies not building AI products are adding AI features to their existing products. Key areas tested: Prompt Engineering: few-shot vs zero-shot prompting, chain-of-thought prompting for reasoning tasks, system prompt design for consistent model behaviour, handling edge cases and out-of-domain queries. Fine-tuning: when to fine-tune vs use in-context learning (ICL). Fine-tuning is appropriate when: you need consistent formatting, you have a specific domain with unique terminology (legal, medical, Indian languages), or you need to reduce prompt length at inference. Parameter-efficient fine-tuning: LoRA (Low-Rank Adaptation) adds trainable rank-decomposition matrices to each layer, allowing fine-tuning with 100x fewer parameters than full fine-tuning. RAG (Retrieval-Augmented Generation): the architecture for giving an LLM access to external knowledge. Components: a document corpus (company documentation, product catalogue, FAQs), an embedding model (converts documents to vectors), a vector database (Pinecone, Chroma, pgvector), and a retriever (finds the most semantically similar documents to the query). Design question: 'Design a RAG system for Flipkart's customer support chatbot that can answer questions about any of 50 million products.' Key considerations: embedding freshness (product descriptions change), retrieval precision (wrong product answers erode trust), hallucination guardrails (the LLM should say 'I don't know' rather than fabricate specs). Multilingual ML for Indian market: Sarvam AI, AI4Bharat, and many Indian companies require engineers who understand multilingual NLP challenges: different scripts (Devanagari, Tamil, Telugu), code-switching (Hinglish), and the limited training data for regional Indian languages compared to English.

Python coding in MLE interviews is different from SWE interviews: the problems are data-centric, not purely algorithmic. Common Python coding tasks in Indian MLE interviews: Implement a confusion matrix from scratch given y_true and y_pred arrays (tests NumPy skills and ML metric understanding). Implement k-means clustering from scratch in NumPy. Write a training loop in PyTorch (dataloader, forward pass, loss computation, backward pass, optimizer step). Implement cross-validation from scratch without sklearn (k-fold split, train on k-1 folds, evaluate on the held-out fold, average metrics). Feature engineering: given a pandas DataFrame of e-commerce transactions, create features for a churn prediction model (recency, frequency, monetary value, time since last purchase, favourite category). These tasks test practical ML engineering skills, not just algorithmic thinking. MLE salary in India 2026: Junior MLE (0-2 years): Rs 15-30 LPA. Mid-level MLE (2-5 years): Rs 30-65 LPA. Senior MLE (5+ years): Rs 60-120 LPA. Principal/Staff MLE: Rs 100-200 LPA. GenAI / LLM engineers command a 20-40% premium over classical ML engineers in 2026 due to the talent shortage. Companies actively competing for LLM talent: Sarvam AI (building Indian language LLMs, strong compensation), Google DeepMind India, Microsoft India, and product companies with major AI feature investments (Myntra AI, Swiggy AI, Flipkart AI). The field is seeing its fastest growth since 2017: candidates with strong Python, PyTorch, and RAG architecture skills are in sustained high demand.