← Uber Interview Insights

Uber·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say
Jun 2026

Summary

ML system design round at Uber for an MLE role. The whole session was basically one big fraud detection question that kept branching into sub-topics. Walked away feeling okay about it but not confident I covered everything they wanted.

Questions Asked (6)

Q1

Design an end-to-end machine learning system for fraud detection on payments or accounts.

System DesignTechnical Trade-offsData Modeling
Author's notes

This was the whole interview basically.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business context—what type of fraud, data volume, latency requirements, and how fraud losses compare to false positive costs. Then walk through the ML lifecycle: data collection and feature engineering, model training with class imbalance handling, real-time serving with low latency, and continuous monitoring with human-in-the-loop feedback. Emphasize trade-offs between precision and recall, model complexity vs. interpretability, and how to handle concept drift in adversarial settings.

Pro tip: Frame the system as a streaming, real-time decision engine with a feedback loop, not just a batch model. Mention Uber's scale and the need for sub-100ms inference, and discuss how you'd handle extreme class imbalance (e.g., 1:1000) with techniques like downsampling, cost-sensitive learning, or anomaly detection.

1. Clarify Requirements and Constraints

Ask about fraud types (payment vs. account takeover), data volume, latency SLAs, and the cost of false positives vs. false negatives. This shapes the entire design.

2. Data Pipeline and Feature Engineering

Design a streaming pipeline (e.g., Kafka) to ingest transactions and user events. Engineer features like transaction velocity, device fingerprint, geolocation, and historical user behavior, ensuring low-latency feature serving.

3. Model Training and Evaluation

Choose models (e.g., gradient boosted trees for tabular data, or deep learning for sequences) and handle class imbalance with resampling or cost-sensitive loss. Use metrics like PR-AUC, recall at fixed precision, and business cost.

4. Real-Time Serving and Decisioning

Deploy model as a microservice with low-latency inference (e.g., <100ms). Implement a rules engine or threshold to convert scores into actions (block, review, allow), and consider model ensembles or cascades.

5. Monitoring, Feedback, and Iteration

Monitor data drift, model performance, and business metrics. Incorporate human review feedback and active learning to retrain models, and set up A/B tests for new models.

Key Points to Mention

  • Class imbalance handling: techniques like SMOTE, undersampling, or focal loss, and why accuracy is misleading.
  • Feature engineering for fraud: velocity checks, time since last transaction, device/IP reputation, and graph-based features for account linkage.
  • Real-time architecture: streaming ingestion, online feature store, low-latency model serving, and fallback rules.
  • Evaluation metrics: precision-recall trade-off, cost-sensitive evaluation, and using business KPIs like fraud loss reduction.
  • Concept drift and adversarial adaptation: continuous monitoring, retraining cadence, and incorporating human feedback.
  • Scalability and reliability: handling Uber-scale traffic, ensuring high availability, and graceful degradation.

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

Q2

How would you choose between decision trees, gradient-boosted trees, and neural networks for a fraud detection model, and what are the tradeoffs?

Technical Trade-offsAlgorithms & Data Structures
Author's notes

Talked through interpretability vs capacity, training cost, and inference latency.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the fraud detection problem's key characteristics: data size, feature types, latency requirements, and class imbalance. Then compare the three model families on those dimensions, highlighting trade-offs in accuracy, interpretability, training/inference speed, and scalability. Conclude with a recommendation that balances business needs (e.g., real-time detection vs. explainability) and suggests a hybrid or ensemble approach if appropriate.

Pro tip: Emphasize that in fraud detection, the cost of false negatives (missed fraud) is often much higher than false positives, so you should tune thresholds and consider asymmetric loss functions. Also mention that gradient-boosted trees often win on tabular data but neural networks can capture complex patterns when you have large-scale, high-cardinality features (e.g., user behavior sequences).

1. Clarify requirements and constraints

Ask about data volume, feature types (tabular vs. sequential), real-time latency, interpretability needs, and class imbalance. This sets the context for model selection.

2. Compare model characteristics

Discuss each model's strengths and weaknesses: decision trees are interpretable but prone to overfitting; gradient-boosted trees excel on tabular data with high accuracy and moderate interpretability; neural networks handle complex, high-dimensional data but require more data and tuning.

3. Evaluate trade-offs in fraud context

Map model traits to fraud detection needs: e.g., gradient-boosted trees for fast inference and strong performance on structured data; neural networks for sequence modeling (e.g., transaction sequences) but with higher latency and lower interpretability.

4. Consider operational factors

Address training time, deployment complexity, monitoring, and maintenance. For example, neural networks may need GPUs and frequent retraining, while tree-based models are easier to deploy and explain to stakeholders.

5. Recommend and justify

Propose a primary choice (e.g., gradient-boosted trees for tabular fraud data) and mention potential ensembles or hybrid approaches. Justify based on the clarified requirements and trade-offs.

Key Points to Mention

  • Class imbalance and evaluation metrics (e.g., precision-recall AUC, cost-sensitive learning)
  • Interpretability vs. accuracy trade-off (e.g., regulatory requirements for explaining decisions)
  • Latency and scalability for real-time fraud detection (e.g., online serving vs. batch)
  • Feature engineering needs: tree models often require less preprocessing, neural networks can learn representations
  • Handling of categorical and high-cardinality features (e.g., embeddings for neural networks, target encoding for trees)
  • Ensemble methods and model stacking to combine strengths

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

Q3

How do you properly evaluate a fraud model given that fraud events are bursty and labels arrive with a delay?

Product Analytics & MetricsTechnical Trade-offs
Author's notes

I knew the time-based train/test split answer cold, so that part went fine.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the challenges of bursty fraud and delayed labels, then propose a robust evaluation framework that combines offline metrics with online monitoring and delayed feedback loops. Emphasize the need for time-aware validation, label maturation analysis, and business-aligned metrics to handle the temporal dynamics.

Pro tip: Use a 'label maturation curve' to determine the optimal evaluation window and avoid premature conclusions; also consider backtesting on historical bursts to simulate real-world performance.

1. Understand the Data Characteristics

Analyze the temporal distribution of fraud events and label delays to quantify burstiness and label arrival patterns. This informs the choice of evaluation metrics and validation strategies.

2. Design Time-Aware Validation

Use time-based splits (e.g., rolling windows) instead of random splits to mimic production. Incorporate label maturation by evaluating only on matured labels or using survival analysis to handle censored labels.

3. Select Appropriate Metrics

Choose metrics that are robust to class imbalance and temporal shifts, such as precision-recall AUC, recall at fixed precision, or cost-based metrics aligned with business impact. Avoid accuracy due to imbalance.

4. Implement Continuous Monitoring

Deploy the model with shadow scoring and monitor performance in real-time using proxy metrics (e.g., alert rates, manual review outcomes) until true labels arrive. Set up alerts for drift in feature and score distributions.

5. Iterate with Delayed Feedback

Once labels mature, backfill evaluation and compare with proxy metrics to calibrate and improve the model. Use this feedback to adjust thresholds and retrain periodically.

Key Points to Mention

  • Time-based validation (e.g., rolling window) to respect temporal order and prevent leakage.
  • Label maturation analysis to determine when enough labels are available for reliable evaluation.
  • Use of precision-recall AUC or recall at high precision due to class imbalance.
  • Proxy metrics for real-time monitoring when labels are delayed (e.g., alert volume, manual review rates).
  • Backtesting on historical bursts to simulate performance during high-fraud periods.
  • Business-aligned cost-sensitive metrics to prioritize fraud detection while minimizing false positives.

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

Q4

What metrics would you use to measure the performance of a fraud detection system, and why?

Product Analytics & MetricsA/B Testing & Experimentation
Author's notes

Said AUC-PR over AUC-ROC under heavy class imbalance and explained why.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem: fraud detection is a highly imbalanced classification problem where the cost of errors is asymmetric. Then, propose a balanced set of metrics that capture both model performance (e.g., precision, recall, AUC-PR) and business impact (e.g., fraud loss prevented, false positive cost). Finally, discuss how to choose and prioritize metrics based on business goals and constraints.

Pro tip: Emphasize that in fraud detection, accuracy is misleading due to class imbalance; instead, focus on precision-recall trade-offs and the cost matrix. Also, mention the importance of monitoring metrics over time and setting up A/B tests to measure the impact of model changes on business KPIs.

1. Understand the problem context

Clarify that fraud detection is an imbalanced classification problem with asymmetric costs: false negatives (missed fraud) are costly, but false positives (blocking legitimate transactions) also harm user experience and revenue.

2. Select model performance metrics

Choose metrics that handle imbalance: precision, recall, F1-score, AUC-ROC, and especially AUC-PR (precision-recall curve). Also consider recall at a fixed precision (e.g., recall@precision=90%) to align with business thresholds.

3. Incorporate business metrics

Translate model performance into business impact: fraud loss prevented, false positive rate (and its cost), customer friction (e.g., declined transactions), and overall net savings. These metrics resonate with stakeholders.

4. Prioritize and balance metrics

Explain how to prioritize based on business goals: if minimizing fraud loss is paramount, focus on recall; if user experience is critical, focus on precision. Use cost-sensitive learning to find the optimal threshold.

5. Monitor and experiment

Discuss the need for continuous monitoring of metrics in production, and using A/B tests to measure the impact of model changes on business KPIs, ensuring that improvements in model metrics translate to business value.

Key Points to Mention

  • Class imbalance and why accuracy is not suitable
  • Precision, recall, F1-score, and AUC-PR
  • Cost-sensitive evaluation and the cost matrix
  • Business metrics: fraud loss prevented, false positive cost, customer friction
  • Threshold tuning based on business constraints
  • A/B testing and monitoring for model performance and business impact

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

Q5

How would you scale this fraud detection system when the data volume exceeds what a single machine can handle, and how do you keep inference latency low?

System DesignTechnical Trade-offs
Author's notes

Distributed training was the easy part.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the scale and latency requirements, then propose a distributed architecture that separates data processing from model serving. Discuss trade-offs between batch and online inference, and how to optimize each component for low latency.

Pro tip: Emphasize that scaling isn't just about adding machines; it's about designing for horizontal scalability, fault tolerance, and efficient data flow. Mention specific technologies like Kafka, Flink, and model serving frameworks (e.g., TensorFlow Serving, Triton) to show practical knowledge.

1. Clarify Requirements

Ask about data volume, velocity, latency SLA, and consistency requirements to scope the problem.

2. Design Distributed Data Pipeline

Propose a scalable data ingestion and processing layer using distributed streaming (e.g., Kafka, Flink) and storage (e.g., HDFS, S3).

3. Scale Model Training and Inference

Discuss distributed training (e.g., Horovod, parameter servers) and model serving with horizontal scaling, caching, and model optimization.

4. Optimize for Low Latency

Detail techniques like model quantization, pruning, hardware acceleration (GPU/TPU), and edge caching to reduce inference time.

5. Monitor and Iterate

Explain how to monitor system performance, detect bottlenecks, and iterate on architecture for continuous improvement.

Key Points to Mention

  • Horizontal scaling with sharding and partitioning
  • Stream processing for real-time feature computation
  • Model serving with load balancing and autoscaling
  • Latency optimization: model compression, hardware acceleration, batching
  • Trade-offs between consistency, availability, and latency (CAP theorem)
  • Fault tolerance and exactly-once processing semantics

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

Q6

How does the model connect to real-time user-facing actions like blocking a transaction or routing to manual review, and how do those outcomes feed back into training?

System DesignCross-functional Alignment
Author's notes

Talked through an orchestration layer that maps model scores to actions like hard block, step-up auth, or review queue.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Walk through the end-to-end ML system: from real-time inference to action execution and feedback loops. Emphasize low-latency serving, decision policies, and how outcomes (e.g., fraud confirmed, manual review results) are logged and used for continuous training. Highlight cross-functional collaboration with product, engineering, and operations teams.

Pro tip: Show awareness of the trade-offs between latency, accuracy, and business impact—e.g., how you balance false positives vs. false negatives in fraud detection, and how you design feedback loops to avoid bias and ensure data quality.

1. Real-time inference and decisioning

Describe how the model is served (e.g., via a low-latency prediction service) and how its output (score/probability) is converted into an action using thresholds or a policy engine.

2. Action execution and integration

Explain how the decision triggers downstream actions: blocking a transaction, routing to manual review, or allowing it. Mention integration with rule engines, workflow systems, and APIs.

3. Outcome capture and labeling

Detail how the system captures the result of the action (e.g., transaction blocked, user complained, manual review confirmed fraud) and how these outcomes become labels for future training.

4. Feedback loop and retraining

Describe the pipeline that aggregates outcomes, handles delayed labels, and periodically retrains the model. Discuss how you mitigate feedback bias (e.g., only seeing outcomes for actions taken).

5. Monitoring and iteration

Mention monitoring for model drift, action distribution, and business metrics. Explain how you close the loop with cross-functional teams to adjust thresholds, policies, and features.

Key Points to Mention

  • Low-latency model serving (e.g., using Michelangelo, Uber's ML platform)
  • Decision policy layer (thresholds, rules, or a bandit approach) to map scores to actions
  • Integration with real-time transaction processing systems (e.g., via Kafka, Flink)
  • Outcome logging and labeling (e.g., chargebacks, manual review results) for supervised learning
  • Handling delayed and biased feedback (e.g., using importance weighting, exploration)
  • Cross-functional collaboration with product, risk, and operations teams to define success metrics and iterate

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