We Are Glad To Inform That You Got The Job- Hope You Hear This Soon-SURVAYU
Home
Posts

INTERVIEW EXPERIENCE OF SWIGGY DATA SCIENTIST ROLE | QUESTIONS & FEEDBACK

 Job Still Open - CLICK HERE for Details

INTERVIEW SLOT DETAILS

  • Candidate Name: Anonymous

  • Slot Time: 29 July 2025, 11:00 AM IST

  • Mode: Online (Google Meet)

  • Interview Level: Entry-Level (0–3 YOE)

  • Role: Data Scientist I

  • Company: Swiggy

🎯 INTERVIEW SIMULATION

⚙️ TECHNICAL ROUND (12 Qs)

Q1. Walk me through how you would build a recommendation system for food delivery ads.

Candidate Answer:
"I’d start by collecting user-item interaction data such as clicks, orders, views. I’d explore collaborative filtering methods like matrix factorization, but also consider deep learning models like neural collaborative filtering. Features like cuisine preference, price sensitivity, and delivery timing will be used. I’d evaluate with precision@k and recall metrics, and experiment using A/B testing for deployment."

Feedback: Strong grasp of both classical and modern techniques. Could have mentioned cold start handling and real-time inference considerations.

Skill Assessed: Recommender Systems, ML modeling
Mistakes to Avoid: Forgetting evaluation metrics or deployment challenges

Q2. How would you optimize ad bidding using ML?

Candidate Answer:
"I’d model bidding as a regression or reinforcement learning problem. First, I’d predict CTR using historical campaign data, then calculate expected revenue per impression and decide optimal bid value. I’d also include constraints like budget pacing and ROI targets."

Feedback: Good conceptual depth. A follow-up on RL implementation or Q-learning would test depth.

Skill Assessed: Ads optimization, ML strategy
Mistakes to Avoid: Giving only textbook-level ideas without metrics or constraints

Q3. Explain how TensorFlow fits into your ML pipeline.

Candidate Answer:
"I use TensorFlow mainly for deep learning models. After feature engineering in Pandas/Spark, I build models in TF using Keras API. I handle training with GPU acceleration, and deploy using TensorFlow Serving for real-time inference."

Feedback: Strong practical flow. Mentioning TF Extended or model versioning would be bonus.

Skill Assessed: TensorFlow proficiency
Mistakes to Avoid: Just theoretical knowledge — TF must be hands-on

Q4. How would you handle a massive dataset (say 10 TB) in your pipeline?

Candidate Answer:
"I’d avoid loading it all into memory. I’d use distributed tools like Spark for ETL, and Parquet or ORC for storage. For ML training, I'd sample intelligently or use mini-batch training with generators. Cloud tools like BigQuery or S3 + Athena could help too."

Feedback: Good awareness of scale-handling tools. Could bring in Data Version Control or Delta Lake for extra maturity.

Skill Assessed: Big Data handling
Mistakes to Avoid: Talking about pandas for big data — no.

Q5. What does “first-principles thinking” mean to you as a data scientist?

Candidate Answer:
"It means breaking down a problem into its most fundamental elements and building the solution from the ground up, rather than relying on analogies or existing solutions."

Feedback: Good — apply it to a real ML case in follow-up.

Skill Assessed: Problem-solving mindset
Mistakes to Avoid: Generic explanation without example

Q6. Write a SQL query to find top 5 most ordered items in last 30 days.

Candidate Answer:

sql
SELECT item_id, COUNT(*) as order_count  
FROM orders  
WHERE order_date >= DATE_SUB(CURRENT_DATE(), INTERVAL 30 DAY)  
GROUP BY item_id  
ORDER BY order_count DESC  
LIMIT 5;

Feedback: Spot on. Efficient and readable.

Skill Assessed: SQL basics
Mistakes to Avoid: Missing the date filter logic or improper GROUP BY

Q7. What’s the difference between recall and precision in ads?

Candidate Answer:
"Precision is how many of the ads shown were actually clicked (relevance), while recall is how many of the relevant ads we successfully showed. For ads, high precision reduces wasted impressions, while high recall ensures we don’t miss good targeting opportunities."

Feedback: Correct. Could add F1 score for balance discussion.

Skill Assessed: Evaluation metrics
Mistakes to Avoid: Mixing up definitions

Q8. How would you use Spark in your data science workflow?

Candidate Answer:
"I’d use PySpark for data cleaning, joining massive datasets, and feature extraction. Spark MLlib can be used for scalable modeling if needed. It’s particularly useful when data doesn’t fit into memory."

Feedback: Solid. Could bring in caching or optimization details.

Skill Assessed: Spark usage
Mistakes to Avoid: Treating Spark like Pandas

Q9. Tell me about a time you took a model from notebook to production.

Candidate Answer:
"In my internship, I built a churn prediction model in Jupyter using scikit-learn. I then containerized it with Docker, used Flask API for serving, and deployed it on AWS EC2 with auto-retraining via cron jobs."

Feedback: Good depth for an early career candidate.

Skill Assessed: ML lifecycle
Mistakes to Avoid: Not addressing deployment or monitoring

Q10. Have you used any GenAI or LLM-based tools in your projects?

Candidate Answer:
"Yes, I built a prompt-based classifier using OpenAI GPT-3 to analyze sentiment in customer reviews. I used LangChain for chaining queries and storing embeddings in Pinecone."

Feedback: Ahead of the curve. Real-world LLM experience is a plus.

Skill Assessed: Generative AI, LLMs
Mistakes to Avoid: Claiming use without real results or evaluation

Q11. How do you keep up with ML research?

Candidate Answer:
"I follow arXiv weekly digests, subscribe to DeepMind and Google AI blogs, and try replicating interesting papers on GitHub."

Feedback: Shows curiosity and learning orientation.

Skill Assessed: Learning & growth
Mistakes to Avoid: Only naming publications — show active engagement

Q12. Give one statistical method you’ve used in ML and why.

Candidate Answer:
"I’ve used logistic regression for binary classification tasks like predicting conversion. It’s interpretable, fast, and a good baseline for many business use cases."

Feedback: Clean and relevant.

Skill Assessed: Applied statistics
Mistakes to Avoid: Not tying method to business goal

💬 BEHAVIORAL ROUND (6 Qs)

Q13. Tell me about a time you disagreed with a teammate on an ML approach.

Q14. How do you explain a complex ML model to a product manager?

Q15. When did you take end-to-end ownership of a data science project?

Q16. How do you handle tight deadlines when a model isn't ready?

Q17. Give an example of collaborating with a cross-functional team.

Q18. How do you prioritize between research ideas and shipping fast?

(Simulated answers & feedback available upon request — want them now or after?)

🧪 SITUATIONAL ROUND (6 Qs)

Q19. Your ads model is underperforming after deployment — how do you debug?

Q20. Your recommender is biased toward expensive items — what would you do?

Q21. The PM wants a model in 1 week but data is messy. How do you proceed?

Q22. A stakeholder wants explainability for a black-box model. What do you do?

Q23. You are given 3 months to improve delivery ETA prediction — how do you plan?

Q24. Model performs well in offline metrics but fails in real-world. Next steps?

🧍‍♀️ HR ROUND (6 Qs)

Q25. Why do you want to work at Swiggy?

Q26. How comfortable are you with remote-first but quarterly travel?

Q27. Are you open to working on weekends if a release is urgent?

Q28. What are your salary expectations?

Q29. What are your career goals for the next 2–3 years?

Q30. Do you have other offers in hand or interview processes?

📝 INTERVIEWER SUMMARY

  • Technical Performance: 8.5/10

  • Communication Skills: 8/10

  • Confidence Level: Medium-High

  • Recommendation: Shortlist for Final Round / Offer Discussion

What the Candidate Missed:

  • More discussion on monitoring ML models post-deployment.

  • Didn’t deeply cover handling data/model drift or continuous learning.

  • Could’ve shown more experience with model explainability (e.g., SHAP, LIME).

Post a Comment