OpenAI India Data Scientist Interview: Rounds, Questions and Preparation

What the job involves

Data Scientists in AI and deeptech companies build and evaluate models, design experiments, and turn messy data into product insights. Day to day you might clean datasets, run A/B tests, prototype ML models, and work with engineers to ship model-driven features. At a company like OpenAI India, the focus is often on large-scale language and multimodal systems, requiring strong statistics, coding, and ML fundamentals.

Role and Technical Knowledge

1. 1

   Explain the bias-variance tradeoff and how you would diagnose it in a model for text classification.

   Define bias and variance, describe underfitting vs overfitting, and mention learning curves or validation error patterns. Relate to text data by discussing overfitting on small datasets or high-dimensional embeddings.

2. 2

   How does a transformer attention mechanism work, and why is it useful for language tasks?

   Cover queries, keys, values, scaled dot-product attention, and multi-head attention. Explain how it captures long-range dependencies better than RNNs and enables parallel training.

3. 3

   What is the difference between precision, recall, and F1 score? When would you prioritize one over another in an AI product?

   Define each metric and give a product example like spam detection (precision) vs medical screening (recall). Mention F1 as a balance and discuss class imbalance.

4. 4

   Describe how you would design an A/B test for a new recommendation model feature.

   State hypothesis, define primary and guardrail metrics, calculate sample size, randomize users, and plan for statistical significance and novelty effects.

Practical Scenarios

1. 5

   You have a dataset with 90% negative and 10% positive labels for a fraud detection model. How would you handle it?

   Suggest resampling (oversample minority, undersample majority), use class weights, choose appropriate metrics like precision-recall AUC, and consider anomaly detection methods.

2. 6

   A model performs well in development but poorly in production. What steps would you take to debug?

   Check data drift, train-serving skew, feature pipeline bugs, and logging. Compare production data distribution to training and validate preprocessing consistency.

3. 7

   How would you explain a complex model's prediction to a non-technical product manager?

   Use SHAP or LIME for local explanations, focus on top features, and avoid jargon. Emphasize actionable insights over technical details.

4. 8

   You need to build a quick prototype for sentiment analysis with limited labeled data. What approach do you take?

   Use a pretrained transformer like BERT, fine-tune on small data, or use zero-shot/few-shot with a large language model. Discuss transfer learning and evaluation.

Behavioural and HR

1. 9

   Tell me about a time you worked on a data project with unclear requirements. How did you proceed?

   Use STAR method: describe the ambiguity, how you asked clarifying questions, made assumptions, and delivered a useful result. Highlight communication.

2. 10

   Why do you want to work in AI and deeptech, and why this company?

   Show genuine interest in AI's potential, mention specific company projects or values, and connect to your skills and long-term goals. Avoid generic answers.

3. 11

   Describe a situation where you had to learn a new tool or technique quickly for a project.

   Pick a concrete example like learning PyTorch or SQL window functions. Explain your learning process, resources used, and the outcome.

4. 12

   How do you handle feedback on your analysis or code that you disagree with?

   Emphasize listening, asking questions, and being open to data-driven arguments. Show you can separate ego from work and iterate.

Coding and Data Manipulation

1. 13

   Write a SQL query to find the second highest salary from an employee table.

   Use a subquery with MAX or ORDER BY with LIMIT/OFFSET. Explain handling ties and NULLs. Show you can think of edge cases.

2. 14

   Given a list of numbers, write a Python function to return the moving average of window size k.

   Use a loop or deque for efficiency. Discuss time complexity and handle edge cases like k larger than list length.

3. 15

   How would you detect and handle missing values in a pandas DataFrame?

   Use isnull(), dropna(), fillna() with mean/median/mode or forward fill. Discuss when to drop vs impute and impact on model.

4. 16

   Write a function to calculate the Euclidean distance between two vectors without using libraries.

   Implement using zip and sum of squared differences, then square root. Mention NumPy alternative for efficiency.