Grab·Machine Learning Engineer·Technical Phone Screen·Intermediate

IntermediatePrefer not to say

Jul 2026Remote

Summary

Grab ML engineer round where they drop a Kaggle link and you basically run a mini competition in real time. The focus isn't really on whether your code compiles, it's more about whether you can think through the problem like an actual ML person.

Questions Asked (2)

Given a dataset of orders, how would you formulate predicting order fulfillment time as an ML problem, and what features would you engineer?

Data ModelingTechnical Trade-offs

Author's notes

This is where most of the conversation lived.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objective and defining the target variable precisely, then outline the ML formulation (regression vs. classification) and discuss feature engineering across temporal, spatial, order, and external data. Emphasize trade-offs like data leakage, real-time vs. batch features, and evaluation metrics aligned with business impact.

Pro tip: Frame the problem in terms of Grab's operational constraints: predicting fulfillment time for real-time dispatch requires low-latency features and careful handling of censored data (e.g., orders not yet delivered). Mention how you'd validate with time-based splits and monitor for concept drift.

1. Clarify Business Objective and Target

Define what 'order fulfillment time' means (e.g., from order placement to delivery) and whether to predict it as a continuous value (regression) or as a category (e.g., on-time vs. late). Consider if the prediction is needed pre-dispatch or post-dispatch.

2. Formulate ML Problem

Choose the ML task: regression for exact time, quantile regression for uncertainty, or classification for SLA breach. Discuss loss functions (e.g., MAE, pinball loss) and evaluation metrics (e.g., RMSE, MAPE, or business KPIs like % on-time).

3. Engineer Features

Brainstorm features across categories: temporal (hour, day, peak), spatial (distance, traffic), order (size, cuisine), driver (experience, load), and external (weather, events). Highlight real-time vs. historical features and avoid leakage.

4. Address Data Challenges

Discuss handling missing data, censored data (orders not yet completed), and imbalanced data if classifying. Mention techniques like survival analysis or using only completed orders with careful bias correction.

5. Validate and Deploy

Propose time-based validation (e.g., rolling window) to mimic production. Discuss deployment considerations: feature store, latency, model retraining frequency, and monitoring for drift.

Key Points to Mention

Target definition: precise start and end points of fulfillment time (e.g., order placement to delivery).
Choice of regression vs. classification and appropriate loss functions (e.g., quantile loss for uncertainty).
Feature engineering: temporal (time of day, day of week), spatial (distance, traffic conditions), order characteristics (size, preparation time), driver attributes (experience, current load), and external factors (weather, events).
Data leakage prevention: ensure features are available at prediction time (e.g., avoid using future information).
Handling censored data: orders in progress may not have final fulfillment time; consider survival analysis or exclusion with bias correction.
Evaluation metrics: align with business goals (e.g., MAE for overall accuracy, % within SLA for operational targets).

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Which model would you choose for this regression task and why, and how would you evaluate it?

Technical Trade-offsProduct Analytics & Metrics

Author's notes

Went with gradient boosted trees, cited interpretability and performance on tabular data with mixed feature types.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the problem context, including data size, feature types, and business objective, then propose a baseline model and a more complex alternative, justifying your choice based on trade-offs. Finally, outline a robust evaluation strategy that includes appropriate metrics, validation techniques, and consideration of business impact.

Pro tip: Demonstrate awareness of Grab's operational constraints (e.g., real-time inference, scalability) by discussing how model choice affects latency and cost, not just accuracy.

1. Clarify the Problem

Ask about data size, feature types, linearity, and business goal to understand the regression task's specifics.

2. Propose Candidate Models

Suggest a baseline (e.g., linear regression) and a more complex model (e.g., gradient boosting), explaining their strengths and weaknesses.

3. Justify Model Choice

Select a model based on trade-offs like interpretability, performance, training time, and inference latency, aligning with business needs.

4. Outline Evaluation Strategy

Describe metrics (e.g., RMSE, MAE), validation (e.g., time-series split), and how to monitor business KPIs.

5. Discuss Deployment and Monitoring

Mention how the model will be deployed, retrained, and monitored for drift and performance degradation.

Key Points to Mention

Trade-offs between model complexity and interpretability
Choice of evaluation metrics (RMSE, MAE, R-squared) and their alignment with business objectives
Cross-validation techniques, especially for time-series or grouped data
Handling of non-linear relationships and feature engineering
Scalability and latency considerations for production
Monitoring and retraining strategies to maintain performance

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.