Aleph Alpha India Data Scientist Interview: Rounds, Questions and Preparation

What the job involves

Data Scientists in AI and deeptech build and evaluate machine learning models, often with a focus on large language models, multimodal systems, and responsible AI. Day to day they clean and analyse datasets, run experiments, track metrics, and collaborate with engineers to ship models into products. At smaller or fast-growing firms, they may also help with data pipelines, evaluation harnesses, and client-facing demos.

Role and technical knowledge

1. 1

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

   Define bias and variance, describe underfitting vs overfitting, and mention learning curves, validation scores, and regularization as diagnostic tools.

2. 2

   How do transformers handle long sequences, and what are the tradeoffs?

   Cover self-attention, quadratic complexity, and techniques like sparse attention, sliding windows, or retrieval to manage long contexts.

3. 3

   What metrics would you use to evaluate a text summarization model, and why?

   Mention ROUGE, BERTScore, and human evaluation, and explain that no single metric captures factual consistency or fluency.

4. 4

   Describe how you would detect and mitigate bias in a model trained on user data.

   Talk about disaggregated evaluation, fairness metrics, data auditing, and mitigation strategies like reweighting or adversarial debiasing.

Practical scenarios

1. 5

   You have a dataset with 90% negative class and 10% positive. How would you build and evaluate a classifier?

   Discuss stratified splits, resampling or class weights, and using precision, recall, F1, and AUC-PR instead of accuracy.

2. 6

   A model performs well offline but poorly in production. How do you debug this?

   Check for data drift, training-serving skew, leakage, and differences in preprocessing; suggest monitoring and shadow deployment.

3. 7

   You need to fine-tune a language model on a small domain dataset. What steps do you take?

   Cover data cleaning, prompt formatting, parameter-efficient fine-tuning like LoRA, validation splits, and overfitting checks.

4. 8

   How would you design an A/B test to measure the impact of a new recommendation model?

   Define hypothesis, random assignment, sample size, primary and guardrail metrics, and duration to avoid peeking.

Behavioural and HR

1. 9

   Why do you want to work in AI and deeptech, and why at a smaller or fast-growing company?

   Show genuine interest in applied research and product impact, and explain comfort with broad ownership and fast iteration.

2. 10

   Tell me about a time you explained a technical result to a non-technical audience.

   Use STAR format, focus on simplifying without losing accuracy, and mention the outcome or decision it enabled.

3. 11

   Describe a project where you had to learn a new tool or method quickly.

   Highlight self-learning, resourcefulness, and how you validated your understanding through a small experiment or prototype.

4. 12

   How do you handle feedback on your analysis or model when a stakeholder disagrees?

   Show openness, ask clarifying questions, re-examine assumptions with data, and focus on shared goals rather than being defensive.

Coding and problem solving

1. 13

   Write a function to compute precision and recall from two lists of binary labels.

   Handle edge cases like zero positives, use clear variable names, and mention that sklearn can do this but you can implement it from scratch.

2. 14

   Given a list of numbers, find the longest increasing subsequence. Explain your approach.

   Discuss dynamic programming O(n^2) and the O(n log n) patience sorting method, and walk through a small example.

3. 15

   How would you use SQL to find the top 3 products by revenue in each category?

   Use window functions like RANK() or DENSE_RANK() partitioned by category, and explain how ties are handled.

4. 16

   Implement a simple train/validation split for a dataset and explain why stratification matters.

   Show code for random and stratified splits, and explain that stratification preserves class proportions, important for imbalanced data.