AI21 Labs India Applied Scientist Interview: Rounds, Questions and Preparation

What the job involves

Applied Scientists at AI-first deeptech companies build and deploy machine learning models that solve real product problems, often working on NLP, generative AI, or recommendation systems. A fresher in this role typically spends time reading papers, implementing prototypes in PyTorch or TensorFlow, running experiments, and collaborating with engineers to ship models to production. The work blends research and engineering, so you need both strong fundamentals and practical coding ability.

Role and technical knowledge

1. 1

   Explain the difference between an Applied Scientist and a Research Scientist in an AI product company.

   Strong answers highlight that Applied Scientists focus on shipping models to production, working with product constraints, and iterating fast, while Research Scientists aim for novel publications and long-term breakthroughs. Mention that both roles overlap in experimentation but differ in deliverables.

2. 2

   How would you evaluate a text summarization model for a product that generates meeting notes?

   Cover both automatic metrics like ROUGE and BERTScore, and human evaluation for fluency, factual consistency, and coverage. Mention the importance of checking for hallucinations and using domain-specific test sets.

3. 3

   What is the attention mechanism in transformers and why is it useful for NLP tasks?

   Explain query, key, and value vectors, scaled dot-product attention, and multi-head attention. Emphasize how it captures long-range dependencies and parallelizes better than RNNs.

4. 4

   Given a dataset with 1 million user reviews, how would you build a sentiment classifier that runs in real time?

   Discuss data cleaning, tokenization, choosing a lightweight model like DistilBERT or a fine-tuned logistic regression on TF-IDF, and optimizing inference with quantization or ONNX. Mention monitoring for drift and retraining.

Practical scenarios and problem solving

1. 5

   Your model performs well on validation but poorly on live traffic. What steps do you take?

   Check for data leakage, distribution shift, and feature pipeline bugs. Compare live vs validation distributions, look at label quality, and consider retraining with recent data or adding drift detection.

2. 6

   You need to reduce inference latency of a BERT-based model by 50 percent without losing more than 2 percent accuracy. How do you approach it?

   Mention knowledge distillation, pruning, quantization, and switching to a smaller model like ALBERT or DistilBERT. Also discuss batching, caching, and hardware acceleration.

3. 7

   How would you handle missing values and class imbalance in a tabular dataset for a fraud detection model?

   For missing values, consider imputation or using models that handle missingness natively. For imbalance, use class weights, resampling, or anomaly detection methods, and evaluate with precision-recall AUC rather than accuracy.

4. 8

   Describe how you would set up an A/B test to measure the impact of a new recommendation model.

   Define a clear metric like click-through rate or dwell time, randomly split users, run for sufficient power, and check for novelty effects. Mention guardrail metrics and statistical significance.

Behavioral and HR questions

1. 9

   Tell me about a time you had to learn a new machine learning concept quickly for a project.

   Use STAR format: describe the project, the gap in your knowledge, the specific steps you took to learn, and the outcome. Show curiosity and ability to self-study.

2. 10

   Why do you want to work at a smaller or fast-growing AI company rather than a large tech firm?

   Focus on impact, ownership, and the chance to work across the stack. Mention that you enjoy fast iteration and seeing your models ship quickly to users.

3. 11

   Describe a situation where your experiment failed. How did you handle it?

   Be honest about the failure, explain what you learned, and how you adjusted your approach. Show resilience and a scientific mindset.

4. 12

   Where do you see yourself in three years as an Applied Scientist?

   Align your growth with the company's needs, such as leading a small project, deepening expertise in NLP or generative AI, and mentoring juniors. Avoid saying you want to leave for a PhD immediately.

Coding and ML implementation

1. 13

   Implement a function to compute the softmax of a vector in Python without using libraries.

   Handle numerical stability by subtracting the max before exponentiation. Write clean code with a loop or list comprehension and test with a simple example.

2. 14

   Write a Python function to calculate the F1 score given true and predicted labels.

   Compute precision and recall from confusion matrix values, then return the harmonic mean. Mention edge cases like zero division.

3. 15

   Given a list of strings, group anagrams together. How would you optimize it?

   Use a dictionary with sorted string as key, or a character count tuple for O(n*k) time. Discuss trade-offs between sorting and counting.

4. 16

   Implement a simple linear regression using gradient descent in NumPy.

   Initialize weights, compute predictions, calculate mean squared error, and update weights with learning rate. Show vectorized operations and mention convergence checks.