← Google Interview Insights

Google·Software Engineer·Onsite - System Design / Architecture·Senior

Senior
Apr 2026

Summary

Google system design round for a software engineer role, focused entirely on building a fraud detection ML system. Pretty deep dive, covered everything from feature engineering to serving infrastructure. Walked out unsure if I'd gone broad enough or just spread myself too thin.

Questions Asked (5)

Q1

Design an end-to-end fraud detection machine learning system for financial transactions.

System DesignTechnical Trade-offsData Modeling
Author's notes

I started with problem framing which felt right, binary classification on transactions with extreme class imbalance.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying requirements and constraints, then design a scalable, real-time system with a feedback loop for continuous learning. Focus on trade-offs between latency, accuracy, and cost, and explain how you would handle imbalanced data and concept drift.

Pro tip: Emphasize the importance of a human-in-the-loop for labeling and model improvement, and discuss how to measure business impact beyond just model metrics (e.g., fraud caught vs. false positives).

1. Clarify Requirements

Ask about scale (transactions per second), latency requirements (real-time vs. batch), data sources, and business goals (e.g., minimize false positives vs. false negatives).

2. High-Level Architecture

Outline components: data ingestion, feature engineering, model training, real-time scoring, and feedback loop. Discuss how to handle streaming data and ensure low-latency predictions.

3. Data and Feature Engineering

Describe how to handle imbalanced data, feature selection (transaction amount, location, time, user history), and techniques like windowed aggregations for real-time features.

4. Model Selection and Training

Discuss model choices (e.g., gradient boosted trees, neural networks), handling concept drift, and retraining strategies. Mention evaluation metrics like precision-recall AUC.

5. Deployment and Monitoring

Explain deployment (e.g., online serving with low latency), monitoring for drift and performance, and incorporating human feedback for continuous improvement.

Key Points to Mention

  • Handling class imbalance (e.g., using SMOTE, class weights, or anomaly detection techniques)
  • Real-time feature computation (e.g., using stream processing frameworks like Apache Flink or Google Cloud Dataflow)
  • Trade-offs between model complexity and latency (e.g., using simpler models for real-time scoring)
  • Concept drift and retraining strategies (e.g., periodic retraining, online learning)
  • Evaluation metrics beyond accuracy (e.g., precision, recall, F1, AUC-PR, and business metrics like fraud detection rate)
  • Scalability and fault tolerance (e.g., using distributed systems, load balancing, and fallback mechanisms)

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

Q2

How would you handle the severe class imbalance between fraudulent and legitimate transactions in your model?

Technical Trade-offsProduct Analytics & Metrics
Author's notes

Rattled off the usual stuff: undersampling, SMOTE, focal loss, reweighting.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging that class imbalance is common in fraud detection and that accuracy is misleading. Then outline a multi-pronged strategy: choose appropriate metrics (e.g., precision-recall AUC), apply data-level techniques (resampling, SMOTE), algorithm-level techniques (class weights, anomaly detection), and consider business costs to set thresholds. Emphasize iterative evaluation and monitoring.

Pro tip: Don't just list techniques; tie them to business impact. For example, explain how you'd optimize for a specific precision at a given recall to balance fraud loss vs. customer friction, and mention that you'd validate with a time-based split to avoid leakage.

1. Define success metrics

Move beyond accuracy to metrics like precision, recall, F1, PR-AUC, and cost-sensitive measures. Align with business goals (e.g., minimize fraud loss while keeping false positives low).

2. Explore data-level techniques

Consider resampling methods: oversampling the minority class (e.g., SMOTE), undersampling the majority, or a combination. Be aware of potential overfitting and validate carefully.

3. Apply algorithm-level techniques

Use class weights, cost-sensitive learning, or anomaly detection algorithms (e.g., isolation forests, autoencoders) that are naturally suited to imbalanced data.

4. Optimize decision threshold

Instead of default 0.5, tune the threshold based on the precision-recall curve to meet business constraints (e.g., max false positive rate).

5. Evaluate and monitor

Use stratified cross-validation and time-based splits. Continuously monitor model performance and retrain as fraud patterns evolve.

Key Points to Mention

  • Precision-Recall AUC is more informative than ROC AUC for imbalanced data.
  • SMOTE and other resampling techniques can help but may introduce synthetic noise.
  • Class weights in algorithms like logistic regression or tree-based models adjust for imbalance.
  • Anomaly detection methods treat fraud as outliers and can be effective when fraud is rare.
  • Threshold tuning based on business costs is crucial; default 0.5 is rarely optimal.
  • Time-based validation prevents data leakage and mimics real-world deployment.

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

Q3

Fraud labels often arrive days later via chargebacks. How does that affect your training pipeline and what do you do about it?

System DesignRoot Cause Analysis
Author's notes

This one tripped me up more than I expected.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Acknowledge that delayed labels create a feedback loop with high latency, requiring a pipeline that handles label delay without degrading model performance. Propose a system that separates immediate feature computation from delayed label ingestion, uses techniques like delayed feedback handling and online learning, and monitors for concept drift. Emphasize the trade-offs between model freshness and accuracy.

Pro tip: Highlight the importance of designing for label delay from the start: use a streaming architecture with a feature store and a label store that can join data asynchronously, and consider using weak supervision or proxy labels to bridge the gap until true labels arrive.

1. Clarify the problem and constraints

Restate the issue: fraud labels arrive days later, causing a delay between prediction and feedback. Discuss the impact on training data freshness, model staleness, and evaluation metrics.

2. Design the data pipeline

Propose a pipeline that ingests transactions in real-time, computes features, and stores them in a feature store. When labels arrive, join them with the stored features to create training examples.

3. Handle delayed labels in training

Use techniques like delayed feedback modeling, importance weighting, or online learning to incorporate labels as they arrive. Consider training on partially labeled data with semi-supervised methods.

4. Monitor and adapt

Set up monitoring for label delay, model performance, and data drift. Implement a retraining schedule that balances freshness with stability, and use A/B testing to validate updates.

5. Evaluate trade-offs and alternatives

Discuss trade-offs between model complexity, latency, and accuracy. Mention alternatives like using proxy labels (e.g., user reports) or ensemble methods to mitigate delay.

Key Points to Mention

  • Feature store for consistent feature computation and storage
  • Delayed label handling techniques: importance weighting, delayed feedback models
  • Online learning or incremental training to incorporate labels as they arrive
  • Monitoring for concept drift and label delay
  • Trade-offs between model freshness and accuracy
  • Use of proxy labels or weak supervision to bridge the gap

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

Q4

How would you serve fraud predictions at low latency in a production environment, and how would you test changes safely?

System DesignA/B Testing & Experimentation
Author's notes

Shadow testing and A/B rollouts, human review queues for borderline cases.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by outlining a low-latency serving architecture for fraud predictions, covering model serving, feature retrieval, and caching. Then discuss safe testing strategies, including shadow deployment, canary releases, and A/B testing with guardrail metrics. Emphasize trade-offs between latency, accuracy, and safety.

Pro tip: Quantify latency budgets (e.g., p99 < 100ms) and explain how you'd monitor and alert on latency regressions. Also, mention the importance of logging prediction context for offline evaluation and debugging.

1. Clarify requirements and constraints

Ask about latency SLAs, throughput, fraud rate, and acceptable false positive/negative rates. Understand data freshness requirements and regulatory constraints.

2. Design low-latency serving architecture

Propose a system with precomputed features, in-memory feature store, model serving via a high-performance framework (e.g., TensorFlow Serving), and caching. Consider edge deployment or regional serving to reduce network latency.

3. Implement safe testing and rollout

Describe shadow mode to test new models without affecting users, canary releases to a small percentage of traffic, and A/B tests with guardrail metrics (e.g., fraud rate, latency). Use feature flags for quick rollback.

4. Monitor and iterate

Set up monitoring for latency, error rates, and model performance. Use online evaluation to detect drift and trigger retraining or rollback. Log predictions for offline analysis.

Key Points to Mention

  • Low-latency serving: model quantization, batching, caching, and async feature retrieval.
  • Feature store with low-latency access (e.g., Redis, Bigtable) and precomputed features.
  • Shadow deployment: run new model in parallel without impacting users, compare predictions.
  • Canary release: gradually shift traffic, monitor key metrics, and rollback if issues.
  • A/B testing: randomize users, define success metrics (e.g., fraud detection rate, latency), and ensure statistical power.
  • Guardrail metrics: latency, error rate, and business metrics to prevent regressions.

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

Q5

A new user or merchant has no transaction history. How does your system handle fraud detection for them?

Technical Trade-offsAdaptability & Ambiguity
Author's notes

Honestly the part I felt best about.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the cold-start problem and the need to balance fraud prevention with user experience. Then describe a multi-layered approach that leverages alternative signals, risk-based authentication, and progressive trust-building. Emphasize monitoring and adaptation as the user generates data.

Pro tip: Mention that you would use a separate model or rule set for new users to avoid false positives, and gradually transition them to the standard model as they build history. This shows you understand the trade-off between security and growth.

1. Acknowledge the cold-start challenge

Explain that without transaction history, traditional fraud models are ineffective, so you need alternative strategies. Highlight the importance of not blocking legitimate new users.

2. Leverage alternative signals

Use device fingerprinting, IP reputation, email/phone verification, and behavioral biometrics to assess risk. These signals can help distinguish between fraudulent and legitimate new users.

3. Apply risk-based authentication and limits

Implement step-up authentication (e.g., OTP, KYC) for higher-risk users and impose transaction limits until trust is established. This balances security with user experience.

4. Monitor and adapt

Continuously monitor new user behavior and update risk scores in real-time. Use feedback loops to improve models and transition users to standard fraud detection as they build history.

Key Points to Mention

  • Cold-start problem in fraud detection
  • Use of alternative data sources (device, IP, behavioral)
  • Risk-based authentication and step-up challenges
  • Progressive trust-building and transaction limits
  • Real-time monitoring and model adaptation
  • Trade-off between fraud prevention and user friction

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