Krutrim Machine Learning Engineer Interview: Rounds, Questions and Preparation

What the job involves

Machine Learning Engineers at AI and deeptech companies like Krutrim build and deploy models that power products such as multilingual chatbots, speech systems, and recommendation engines. Day to day, you will clean and preprocess large datasets, train and fine-tune models (often transformer-based), run experiments, and optimise models for latency and cost on cloud or edge hardware. You will also collaborate with data engineers, product managers, and backend teams to ship models into production and monitor their performance.

Role and technical knowledge

1. 1

   Explain the difference between bagging and boosting. Give an example of when you would use each for a tabular dataset in a production ML pipeline.

   Strong answer contrasts variance reduction (bagging, e.g., Random Forest) vs bias reduction (boosting, e.g., XGBoost), and mentions trade-offs in training time, interpretability, and overfitting risk.

2. 2

   How does the attention mechanism work in a transformer? Why is it useful for multilingual NLP tasks like those in Indian languages?

   Cover query-key-value computation, scaled dot-product, and multi-head attention. Link to handling long-range dependencies and language-agnostic representations.

3. 3

   You have a dataset with 90% negative and 10% positive labels. Which metrics would you use and why? How would you handle the imbalance?

   Mention precision, recall, F1, and AUC-ROC instead of accuracy. Discuss resampling, class weights, and threshold tuning.

4. 4

   What is the difference between model parameters and hyperparameters? How do you tune hyperparameters efficiently for a deep learning model?

   Define both clearly. Mention grid search, random search, Bayesian optimisation, and early stopping with a validation set.

Practical scenarios

1. 5

   Your model performs well on the test set but poorly in production. What could be the reasons and how would you debug it?

   Discuss data drift, leakage, different preprocessing in production, and lack of monitoring. Suggest logging inputs, comparing distributions, and retraining.

2. 6

   You need to deploy a large language model for a low-latency chatbot on a limited GPU budget. What techniques would you use?

   Mention quantization, pruning, knowledge distillation, caching, and using smaller fine-tuned models. Also talk about batching and serving frameworks like TorchServe or Triton.

3. 7

   How would you build a named entity recognition system for Indian language text with limited labelled data?

   Suggest transfer learning from multilingual models (e.g., mBERT, XLM-R), data augmentation, active learning, and using rule-based or gazetteer features.

4. 8

   A model you trained gives inconsistent outputs for similar inputs. How do you make it more robust?

   Cover data augmentation, adversarial training, ensembling, and checking for numerical instability. Mention testing with edge cases.

Behavioural and HR

1. 9

   Tell me about a machine learning project you built from scratch. What was your role and what did you learn?

   Use STAR format. Highlight problem, data, model choice, evaluation, and a specific lesson about deployment or iteration.

2. 10

   Describe a time you had to explain a technical concept to a non-technical teammate. How did you ensure understanding?

   Show empathy and clarity. Mention analogies, visual aids, and checking for questions. Link to working with product or design teams.

3. 11

   Why do you want to work in AI and deeptech, specifically at a company building foundational models for India?

   Connect personal motivation to impact at scale, multilingual challenges, and local data. Avoid generic praise; mention specific technical interests.

4. 12

   You are given a deadline that seems impossible. How do you prioritise and communicate?

   Discuss breaking down tasks, identifying must-haves vs nice-to-haves, raising blockers early, and proposing a reduced scope with clear trade-offs.

Coding and implementation

1. 13

   Implement k-means clustering from scratch in Python using only NumPy. How would you choose k?

   Write clear loops for assignment and update steps. Discuss elbow method, silhouette score, and limitations like sensitivity to initialisation.

2. 14

   Given a list of sentences, write a function to compute the cosine similarity between all pairs efficiently. What data structure would you use?

   Show vectorisation with NumPy or PyTorch. Mention normalising vectors and using matrix multiplication. Discuss memory for large N.

3. 15

   Write a PyTorch training loop for a simple feedforward network on a classification task. Include loss, optimiser, and validation.

   Include zero_grad, forward, loss, backward, step. Add model.train() and model.eval() with torch.no_grad() for validation. Mention tracking metrics.

4. 16

   How would you split a large CSV file into train, validation, and test sets while ensuring no data leakage across time?

   For time-series, use chronological splits. Explain why random split is wrong for temporal data and how to use scikit-learn's TimeSeriesSplit.