Zomato Machine Learning Engineer Interview: Questions, Experience & Prep (2026)

Zomato is one of India's most recognizable food-tech companies, processing orders at high volume across many cities. The ML team powers everything from restaurant recommendations and delivery time estimates to fraud detection and demand forecasting. As of July 2026, knok jobradar shows 9 open Machine Learning Engineer roles at Zomato, signaling active hiring across the ML org.

Candidates typically go through multiple rounds covering machine learning fundamentals, system design for large-scale ML, and at least one round focused on past experience and problem-solving approach. The process is known for being applied and product-focused: Zomato interviewers want to see that you can connect ML techniques to real business outcomes, not just recite theory.

1. Restaurant recommendation system: How would you design a personalized restaurant recommendation system for the Zomato home feed?

1. ETA prediction: Walk me through how you would build a model to predict delivery time for an order, from placement to doorstep.

1. Cold start problem: A new restaurant joins Zomato and has zero order history. How do you handle recommendations or ranking for it?

1. Fraud detection: How would you detect fraudulent orders or fake restaurant reviews at scale?

1. Search ranking: How would you rank search results when a user types 'biryani' in the Zomato search bar?

1. Demand forecasting: How would you build a model to forecast how many delivery partners are needed in a specific zone at a given hour?

1. Class imbalance: Fraud events are rare compared to genuine orders. How do you handle severe class imbalance in your model?

1. A/B testing: You have a new recommendation algorithm. How do you run an A/B test and decide whether to ship it?

1. Feature engineering: What features would you extract from GPS trace data to improve delivery time prediction?

1. Model monitoring: Once your model is in production, how do you detect and respond to data drift or performance degradation?

1. Prep time prediction: How would you predict how long a restaurant will take to prepare an order, and how does this feed into the overall delivery ETA?

1. Trade-off question: Your recommendation model improves click-through rate but reduces average order value. What do you do?

Q: How would you design a personalized restaurant recommendation system for the Zomato home feed?

*Situation:* In a past role, I worked on a content recommendation system for an e-commerce platform with a large product catalog and high daily traffic.

*Task:* I was asked to improve personalization on the home page, which was showing mostly popular items rather than items relevant to each individual user.

*Action:* I built a two-stage retrieval-and-ranking pipeline. The first stage used collaborative filtering to retrieve a candidate set of items a user was likely to engage with, based on behavior from similar users. The second stage applied a gradient boosted ranking model that scored each candidate using user-level features (past orders, cuisine preferences, time of day), item-level features (ratings, distance, price band), and context features (device, session intent). I also handled the cold start case by falling back to content-based signals for new users.

*Result:* Click-through rate on the home feed improved in A/B testing, and we saw a sustained lift in repeat purchases over the following quarter. I would apply the same architecture at Zomato, adding delivery time and restaurant prep time as additional ranking signals unique to food delivery.

---

Q: How would you handle fraud detection for fake restaurant reviews at scale?

*Situation:* At a previous company, we discovered a pattern of fake reviews being posted to manipulate product rankings.

*Task:* I was responsible for building a real-time system to detect and flag suspicious reviews before they affected rankings.

*Action:* I designed a feature set around behavioral signals: account age, posting velocity, linguistic patterns (sentence structure similarity across reviews from the same IP range), and device fingerprints. I trained a gradient boosted classifier on labeled data from a prior manual review campaign. Because fraudulent reviews were far rarer than genuine ones, I used a combination of oversampling and cost-sensitive learning to handle the class imbalance. I also built a rule-based pre-filter to catch obvious bot patterns before the model even ran.

*Result:* Precision on the held-out test set was high enough to route flagged reviews to a human review queue rather than auto-removing them, keeping false positive rates low. Fraudulent review volume on the platform dropped significantly in the months following launch.

---

Q: Describe how you would monitor a model in production and respond to drift.

*Situation:* A demand forecasting model I had trained started showing degraded accuracy about three months after deployment, without any obvious trigger.

*Task:* I needed to diagnose the root cause and restore model performance without a full re-training cycle if possible.

*Action:* I set up feature distribution monitoring using population stability index checks that ran nightly. I found that one input feature (historical average order volume per zone) had shifted significantly because a competitor had exited certain markets, changing baseline demand patterns. I retrained the model on a more recent data window, added a feature capturing relative market share signals, and set up automated alerts for PSI breaches above a threshold defined from historical variance.

*Result:* Forecast accuracy returned to baseline within two weeks of the updated model going live. The alerting system caught two subsequent drift events in the following year, both addressed before they impacted downstream planning.

Use STAR for experience questions. Every behavioral question at Zomato (and most Indian tech companies) expects a Situation, Task, Action, Result structure. Keep Situation and Task brief (two to three sentences combined) and spend most of your time on Action and Result.

Use a structured ML design framework for system design questions. Candidates report that Zomato interviewers respond well to answers that follow a clear sequence:

1. Clarify the business objective and how you will measure success (metric alignment first).
2. Define the ML problem type (ranking, classification, regression, etc.) and justify the framing.
3. Describe data sources and feature engineering, with specific examples relevant to food delivery.
4. Choose a model family and explain the trade-offs (interpretability vs. accuracy, latency vs. quality).
5. Explain how you would train, evaluate, and deploy the model.
6. Address monitoring, retraining triggers, and failure modes.

For trade-off questions, name the competing objectives explicitly, propose a combined metric or north star, and explain which levers you would pull and in what order. Interviewers want to see that you can reason under ambiguity, not just pick one side.

For coding and algorithm questions, speak your thinking out loud before writing any code. Zomato interviewers typically value clear problem decomposition over a fast but silent solution.

Applied ML thinking over pure theory. Zomato operates at scale in a messy real-world environment: GPS noise, variable restaurant behavior, and seasonal demand spikes. Interviewers want to see that you connect every ML choice back to a concrete business or product outcome.

Familiarity with food-tech or marketplace ML problems. You do not need prior food-tech experience, but showing that you have thought about recommendation systems, ETA prediction, or demand forecasting will set you apart from candidates who only speak in generic ML terms.

Comfort with large-scale systems. Candidates report that questions about data pipelines, feature stores, and low-latency serving come up frequently. Know the difference between batch and real-time inference and when each is appropriate.

Strong communication. Interviewers at product companies like Zomato typically want to follow your reasoning step by step. Jumping to a solution without explaining your assumptions can signal poor collaborative instincts.

Ownership mindset. Behavioral questions tend to probe whether you take end-to-end responsibility for a model, from data collection through monitoring, rather than handing off at deployment.

Week 1: ML foundations and Zomato domain.
Review core concepts: gradient boosting, neural collaborative filtering, learning to rank, and time-series forecasting. Then spend time reading the Zomato engineering blog (search 'Zomato engineering blog' to find the official source) to understand the actual problems their ML team has written about publicly.

Week 2: System design practice.
Practice designing end-to-end ML systems for at least three Zomato-relevant problems: a recommendation system, a delivery ETA predictor, and a fraud detector. Aim for two structured walkthroughs per day using the six-step framework described in the Answer Frameworks section.

Week 3: Coding and statistics.
Refresh your Python and SQL skills. Practice writing clean pandas and scikit-learn code under time pressure. Review statistical concepts that come up in ML interviews: hypothesis testing for A/B tests, precision-recall trade-offs, and cross-validation strategies.

Week 4: Mock interviews and behavioral prep.
Do at least four mock interviews with a peer or on a platform that supports ML interview practice. Prepare five to seven STAR stories covering model improvement, handling ambiguity, cross-functional collaboration, and a project that did not go as planned. Candidates report that Zomato behavioral rounds are thorough.

Throughout: Track your applications carefully. knok checks 150+ job sites nightly, applies to roles matching your resume, and messages HR on your behalf, so you can focus your energy on preparation rather than job hunting logistics.

Jumping to modeling before defining the problem. A very common mistake in ML design rounds is proposing a specific algorithm (for example, a neural network) before establishing what metric you are optimizing, what data is available, and what the latency constraints are. Interviewers at product companies penalize this heavily.

Ignoring cold start and edge cases. Food delivery platforms constantly onboard new restaurants and new users. If your recommendation or ranking system has no answer for cold start, interviewers will push back. Always address edge cases proactively.

Treating class imbalance as an afterthought. Fraud detection, churn prediction, and anomaly detection all involve rare positive classes. Candidates who only mention 'use SMOTE' without discussing why or what trade-offs it introduces signal shallow experience.

Generic STAR answers. Saying 'I improved model accuracy' without specifying what changed, why it was hard, and what the business impact was is a weak answer. Tie every result back to a metric or a business outcome.

Not asking clarifying questions in system design. Walking straight into a solution without clarifying scope, scale, and success criteria is a common mistake. The constraints change the right answer significantly at the scale Zomato operates.

Skipping monitoring and retraining. Many candidates design a model, evaluate it offline, and stop there. A complete answer covers how the model behaves in production over time, including drift detection and retraining triggers.