Jasper AI India Applied Scientist Interview: Rounds, Questions and Preparation

What the job involves

Applied Scientists at AI and deeptech companies design, train, and evaluate machine learning models that power generative AI products, often focusing on large language models, retrieval, and personalization. Day to day, they run experiments, analyze results, and collaborate with engineers and product managers to ship model improvements. Freshers typically start by owning smaller model components or evaluation pipelines.

Role and Technical Knowledge

1. 1

   Explain how a transformer model generates text, from input tokens to output probabilities.

   Cover tokenization, positional encodings, self-attention, feed-forward layers, and softmax over vocabulary. Mention autoregressive decoding.

2. 2

   What is the difference between fine-tuning and prompt engineering for a generative AI product?

   Discuss cost, data needs, latency, and control. Fine-tuning changes weights; prompting steers without weight updates. Give a use case for each.

3. 3

   How would you evaluate the quality of a summarization model for a marketing copy tool?

   Mention both automatic metrics like ROUGE and human evaluation for fluency, factuality, and brand tone. Discuss A/B testing with real users.

4. 4

   Describe the bias-variance tradeoff and how it applies to large language models.

   Explain underfitting vs overfitting, and how model size, data, and regularization affect it. For LLMs, discuss memorization vs generalization.

Practical Scenarios

1. 5

   You are building a RAG system for a customer support chatbot. How do you choose and evaluate the retriever?

   Discuss embedding models, vector databases, recall@k, and end-to-end answer accuracy. Mention chunking strategies and reranking.

2. 6

   Your model's output suddenly becomes repetitive and less diverse after a fine-tuning run. What do you check?

   Check learning rate, overfitting, data quality, and decoding parameters like temperature and top-p. Suggest validation on held-out prompts.

3. 7

   How would you reduce inference cost for a large model without hurting quality much?

   Mention quantization, distillation, pruning, caching, and batching. Discuss tradeoffs and when to use each.

4. 8

   A product manager asks for a feature that generates ad headlines. Outline your approach from data collection to deployment.

   Cover data sourcing, prompt design or fine-tuning, evaluation with human raters, safety checks, and deployment with monitoring.

Behavioural and HR

1. 9

   Tell me about a time you had to learn a new ML technique quickly for a project.

   Use STAR format. Highlight self-learning, applying it to a real problem, and the outcome. Show curiosity and execution.

2. 10

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

   Focus on analysis, learning from negative results, and iterating. Show resilience and scientific thinking.

3. 11

   Why do you want to work as an Applied Scientist in generative AI, and why our company?

   Connect your skills to the company's products and mission. Show awareness of the deeptech sector and a specific interest in applied research.

4. 12

   How do you prioritize when you have multiple experiments to run and limited compute?

   Discuss impact vs effort, quick wins, and communication with stakeholders. Mention tracking and reproducibility.

Coding and Math

1. 13

   Implement a function to compute the attention scores for a single query vector against a set of key vectors.

   Write clear code for dot product, scaling, and softmax. Discuss numerical stability and vectorization.

2. 14

   Given a dataset with missing values and outliers, how would you preprocess it for a neural network?

   Mention imputation, normalization, outlier detection, and train/validation split. Discuss why each step matters.

3. 15

   Explain the central limit theorem and why it matters for A/B testing in ML products.

   Describe sampling distributions and confidence intervals. Connect to estimating treatment effects and making decisions from noisy data.

4. 16

   Write a Python function to calculate the perplexity of a language model given a list of token probabilities.

   Show formula: perplexity = exp(-average log probability). Handle edge cases like zero probabilities with smoothing.