← Credit Karma Interview Insights

Credit Karma·Machine Learning Engineer·Onsite - System Design / Architecture·Senior

Senior
Apr 2026

Summary

ML system design loop for Credit Karma, focused entirely on recommendation and ranking for their financial product marketplace. Heavy on the ML infra side, lots of follow-ups about label sparsity and multi-stage architectures. Felt more like a research discussion than a standard interview.

Questions Asked (10)

Q1

How would you design an ML system for Credit Karma to maximize revenue from financial product recommendations?

System DesignPricing & MonetizationProduct Strategy
Author's notes

Big open-ended opener.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objective—maximizing revenue from financial product recommendations—and the constraints (e.g., user trust, regulatory compliance). Then outline a system that balances user relevance with revenue optimization, covering data, modeling, serving, and evaluation.

Pro tip: Emphasize that revenue maximization must be balanced with long-term user trust and regulatory compliance; propose guardrail metrics like user satisfaction and retention to avoid short-term gains that harm the platform.

1. Clarify Business Goals and Constraints

Define what 'maximize revenue' means: is it total revenue, revenue per user, or long-term value? Identify constraints like user experience, regulatory rules, and fairness.

2. Data and Feature Engineering

Identify data sources: user financial profiles, product catalogs, historical interactions, and revenue data. Engineer features like user affinity, product profitability, and contextual signals.

3. Modeling Approach

Design a two-stage system: candidate generation (e.g., collaborative filtering) and ranking (e.g., a multi-objective model that predicts click-through rate and revenue per conversion). Consider using a revenue-weighted objective.

4. Serving and Experimentation

Deploy with an A/B testing framework to measure revenue impact. Use online learning to adapt to changing user behavior and product offerings.

5. Evaluation and Guardrails

Define offline metrics (e.g., AUC, revenue lift) and online metrics (e.g., revenue per user, conversion rate). Monitor guardrail metrics like user satisfaction and retention to ensure long-term health.

Key Points to Mention

  • Multi-objective optimization balancing user relevance and revenue
  • Revenue-weighted ranking or bid-based auction mechanisms
  • Exploration vs exploitation to avoid feedback loops
  • Regulatory compliance and fairness in financial recommendations
  • Long-term user value vs short-term revenue
  • A/B testing and causal inference for measuring revenue impact

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

Q2

Should you use a single model across the entire funnel or separate models for each stage like CTR, application rate, and approval rate?

System DesignTechnical Trade-offs
Author's notes

I argued for separate models first, then they pushed back asking about maintenance overhead and feature overlap.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging that both approaches have trade-offs and the right choice depends on data availability, business goals, and model complexity. Then, discuss the benefits of a unified model (e.g., shared representations, data efficiency) versus stage-specific models (e.g., tailored features, interpretability). Finally, propose a hybrid or staged approach, emphasizing the importance of experimentation and monitoring.

Pro tip: Emphasize that in production, you often need to balance model performance with maintainability and latency; a single model can reduce operational overhead but may sacrifice stage-specific accuracy. Mention that at Credit Karma, where the funnel involves credit decisions, interpretability and compliance might favor separate models for approval rate.

1. Clarify the funnel and objectives

Define each stage (CTR, application rate, approval rate) and the business metrics they impact. Understand whether stages share common signals or have distinct drivers.

2. Evaluate data and feature availability

Assess if you have enough data per stage to train separate models. Consider whether features are shared across stages or unique to each.

3. Compare modeling approaches

Discuss pros and cons of a single multi-task model (e.g., shared layers, multi-task learning) versus separate models (e.g., specialized features, independent optimization).

4. Consider operational constraints

Factor in latency, maintainability, monitoring, and compliance requirements. A single model may simplify deployment but complicate debugging.

5. Recommend and justify

Propose a solution (e.g., start with separate models, then explore multi-task learning) and suggest A/B testing to validate performance.

Key Points to Mention

  • Multi-task learning can leverage shared representations and improve data efficiency for related tasks.
  • Separate models allow for tailored feature engineering and hyperparameter tuning per stage.
  • Data sparsity in later funnel stages (e.g., approval rate) may necessitate separate modeling or transfer learning.
  • Operational complexity: a single model reduces deployment and maintenance overhead but may be harder to interpret.
  • Business context: Credit Karma's regulatory environment may require explainability, favoring separate models for approval decisions.
  • Experimentation: A/B testing is crucial to compare unified vs. separate models in terms of business metrics.

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

Q3

How do you handle conversion labels that arrive days or weeks after the original impression?

System DesignTechnical Trade-offs
Author's notes

Talked about join windows and cutting off labels at a fixed delay, but honestly I fumbled the part about how this creates bias in your training distribution.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the delayed feedback problem and its impact on model training and evaluation. Then describe a robust system design that handles delayed labels, including data pipelines, model retraining strategies, and evaluation metrics. Emphasize trade-offs between freshness and accuracy, and how you would validate the approach.

Pro tip: Mention the importance of logging the timestamp of label arrival and using it to weight samples or adjust for delay bias. Also, discuss how you would simulate delayed labels in offline experiments to tune the system.

1. Clarify the problem and constraints

Ask about the expected delay distribution, volume of data, and business impact. Understand if real-time predictions are needed or if batch processing suffices.

2. Design data pipeline for delayed labels

Propose a system that ingests impressions immediately and joins labels when they arrive, using a mutable store or streaming architecture. Ensure data versioning and time-travel capabilities.

3. Model training and retraining strategy

Discuss using delayed labels in training with techniques like importance weighting, survival analysis, or delayed feedback models. Decide on retraining frequency and how to handle missing labels.

4. Evaluation and monitoring

Define metrics that account for delay, such as time-windowed AUC or calibration. Set up monitoring for label arrival rates and model performance drift.

5. Trade-offs and iteration

Highlight trade-offs between model freshness and accuracy, complexity of implementation, and business impact. Suggest A/B testing to validate the approach.

Key Points to Mention

  • Delayed feedback problem in conversion modeling
  • Importance weighting or sample weighting based on delay
  • Survival analysis or hazard modeling for time-to-conversion
  • Streaming vs batch processing architectures
  • Time-based validation and avoiding data leakage
  • Business impact of delayed labels on bidding or ranking

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

Q4

Conversion rates in this domain are extremely sparse, around 0.1 to 0.5 percent. How do you deal with that during training?

System DesignTechnical Trade-offs
Author's notes

Negative downsampling was the obvious answer and I gave it immediately.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging the challenge of extreme class imbalance and its impact on training. Then, outline a multi-faceted strategy that includes data-level techniques (e.g., resampling, augmentation), algorithmic adjustments (e.g., cost-sensitive learning, specialized loss functions), and evaluation metrics (e.g., PR-AUC, recall@k). Emphasize the importance of aligning the approach with business objectives and validating with proper offline and online testing.

Pro tip: Don't just list techniques—explain the trade-offs and how you would prioritize them based on the specific business context, such as the cost of false positives vs. false negatives. Also, mention that you would start with a simple baseline and iterate, rather than over-engineering from the start.

1. Understand the Business Context and Define Success

Clarify the business goal and the cost associated with different types of errors. Determine what metric matters most (e.g., precision at a fixed recall, lift, or expected value).

2. Data-Level Techniques

Discuss resampling methods like oversampling the minority class (with SMOTE or variants) or undersampling the majority class. Consider data augmentation if applicable, and ensure proper cross-validation to avoid leakage.

3. Algorithmic and Loss Function Adjustments

Mention cost-sensitive learning, class weights, and specialized loss functions like focal loss that focus on hard examples. Also, consider using tree-based models that handle imbalance well or anomaly detection approaches.

4. Evaluation and Validation

Use appropriate metrics like PR-AUC, recall at high precision, or F-beta score. Ensure validation sets reflect the real-world distribution and consider time-based splitting if temporal patterns exist.

5. Iterate and Monitor in Production

Start with a simple model and iterate. Monitor performance in production, as conversion rates may drift, and be prepared to retrain or adjust thresholds dynamically.

Key Points to Mention

  • Class imbalance handling: oversampling, undersampling, SMOTE, and their trade-offs.
  • Cost-sensitive learning and class weights to penalize misclassification of the minority class.
  • Specialized loss functions like focal loss to focus on hard-to-classify examples.
  • Evaluation metrics: PR-AUC, recall@k, F-beta, and lift, instead of accuracy.
  • Proper cross-validation and avoiding data leakage when resampling.
  • Business alignment: optimizing for expected value and considering the cost of false positives/negatives.

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

Q5

If you apply negative sampling during training, how do you correct for that bias at inference time?

System DesignAlgorithms & Data Structures
Author's notes

This one actually went well.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Explain that negative sampling introduces a bias because the model is trained to distinguish true positives from a sampled subset of negatives, not the full distribution. To correct at inference, you need to adjust the scores using the sampling probability, typically by subtracting log(q(x)) from the logit, where q(x) is the negative sampling distribution. Then discuss practical implications and alternatives.

Pro tip: Mention that in practice, many systems avoid explicit correction by using negative sampling only for training efficiency and then calibrating scores on a validation set or using a two-tower architecture where the full softmax is approximated. This shows awareness of real-world trade-offs.

1. Acknowledge the bias

State that negative sampling biases the model because it only sees a subset of negatives, leading to overestimation of positive class probability relative to the true distribution.

2. Derive the correction

Show that the corrected probability is proportional to the model output divided by the negative sampling probability. In log-odds, subtract log(q(x)) from the logit.

3. Apply at inference

At inference, for each item, compute the adjusted score: score(x) - log(q(x)), where q(x) is the probability of sampling x as a negative. This yields unbiased ranking or probability estimates.

4. Discuss practical considerations

Note that q(x) must be known and stable; if it's uniform, correction is a constant shift. Also mention that for ranking, monotonic transformations may not affect order, but for calibrated probabilities, correction is essential.

5. Alternative approaches

If correction is difficult, consider using full softmax at inference (if feasible) or calibrating on a held-out set. Mention that some systems ignore correction if only ranking is needed.

Key Points to Mention

  • Negative sampling changes the effective prior of negatives, so the model's output is not calibrated to the true distribution.
  • The correction formula: p_true(x) ∝ p_model(x) / q(x), where q(x) is the negative sampling probability.
  • In logit space, the correction is a simple subtraction: logit_corrected = logit_model - log(q(x)).
  • If q(x) is uniform, the correction is a constant offset, which does not affect ranking but affects probability calibration.
  • For ranking tasks, correction may be unnecessary if the sampling distribution is independent of the query, but for probability estimates it is crucial.
  • Practical systems often use negative sampling for training efficiency and then calibrate scores on a validation set or use a full softmax at inference if the item set is small.

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

Q6

You're training on multiple objectives like CTR, conversion, and approval rate. How do you combine those into a final ranking score?

System DesignPricing & MonetizationTechnical Trade-offs
Author's notes

Talked through weighted combinations and constrained optimization.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business goal and how each objective maps to it, then describe a multi-objective optimization framework that combines objectives into a single score. Explain how you would weight or learn the combination, and how you would validate and iterate.

Pro tip: Emphasize that the combination should be dynamic and context-aware, and that you'd use online experiments to tune weights rather than relying solely on offline metrics.

1. Clarify Business Objectives

Understand how each objective (CTR, conversion, approval rate) contributes to the overall business goal, such as revenue or user satisfaction. Identify any constraints or trade-offs.

2. Choose a Combination Method

Select a method to combine objectives, such as weighted sum, product, or learning-to-rank with multi-task learning. Consider linear vs. non-linear combinations.

3. Determine Weights or Parameters

Decide how to set weights: manually via business input, or automatically via optimization (e.g., Bayesian optimization, reinforcement learning). Ensure weights reflect current priorities.

4. Validate and Iterate

Use offline evaluation (e.g., NDCG, precision@k) and online A/B tests to measure impact. Iterate on the combination strategy based on results.

5. Monitor and Adapt

Continuously monitor performance and adapt weights or model as business goals evolve. Implement guardrail metrics to prevent degradation in any objective.

Key Points to Mention

  • Multi-task learning or multi-objective optimization techniques
  • Weighted sum vs. other combination functions (e.g., product, rank aggregation)
  • Normalization of objectives to comparable scales
  • Business alignment and stakeholder input for weights
  • Online experimentation (A/B testing) for tuning and validation
  • Trade-offs and guardrail metrics to avoid sacrificing one objective

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

Q7

What ranking objective function would you actually use to combine multi-objective scores into a final ordering?

System DesignTechnical Trade-offs
Author's notes

Felt like a follow-up to the previous question but more concrete.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the business objectives and constraints, then propose a weighted linear combination as a baseline, and discuss more advanced methods like learning-to-rank or multi-objective optimization. Emphasize that the choice depends on whether you have labeled data, the need for interpretability, and the trade-offs between objectives.

Pro tip: Mention that in production, you often start with a simple weighted sum and then iterate using online metrics and A/B tests, rather than over-engineering upfront. Also, highlight the importance of calibrating scores before combining them.

1. Clarify Objectives and Constraints

Ask about the specific business goals, such as maximizing user engagement while minimizing risk, and any regulatory constraints. Understand what data is available and whether there are labeled outcomes.

2. Choose a Baseline Approach

Propose a weighted linear combination of normalized scores as a simple, interpretable baseline. Explain how weights can be set via business rules or learned from data.

3. Consider Advanced Methods

If data and complexity allow, discuss learning-to-rank algorithms (e.g., LambdaMART) or multi-objective optimization techniques like Pareto optimization. Mention the trade-offs in terms of interpretability and development effort.

4. Address Calibration and Normalization

Explain the need to calibrate scores to a common scale (e.g., via Platt scaling or isotonic regression) before combining, to ensure fair weighting.

5. Evaluate and Iterate

Describe how to evaluate the ranking using offline metrics (e.g., NDCG) and online A/B tests, and how to iterate on the objective function based on results.

Key Points to Mention

  • Weighted linear combination as a baseline
  • Learning-to-rank methods (e.g., LambdaMART, RankNet)
  • Multi-objective optimization and Pareto frontier
  • Score calibration and normalization techniques
  • Offline evaluation metrics like NDCG and online A/B testing
  • Trade-offs between interpretability, complexity, and performance

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

Q8

Compare a two-stage retrieval plus reranking architecture to scoring all candidate items directly in a single pass.

System DesignTechnical Trade-offs
Author's notes

Pretty standard for anyone who's done recommender work.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the problem context, such as dataset size, latency requirements, and quality metrics. Then compare the two architectures across key dimensions like computational cost, latency, accuracy, and scalability. Conclude with a recommendation based on trade-offs and mention hybrid approaches if relevant.

Pro tip: Emphasize that the choice depends on the scale and latency constraints; for large-scale systems, two-stage is often necessary, but for smaller or latency-critical systems, single-pass can be simpler and effective. Also, mention that reranking can be seen as a form of learned re-scoring that can incorporate more features.

1. Clarify Requirements

Ask about the scale of the candidate set, latency constraints, and quality metrics to understand the problem context.

2. Describe Two-Stage Architecture

Explain that it first retrieves a subset of candidates using a fast, recall-oriented model, then reranks them with a more complex, precision-oriented model.

3. Describe Single-Pass Architecture

Explain that it scores all candidates directly with a single model, which must balance speed and accuracy.

4. Compare Trade-offs

Discuss computational cost, latency, accuracy, scalability, and complexity for both approaches.

5. Recommend and Conclude

Provide a recommendation based on the requirements and mention potential hybrid approaches or optimizations.

Key Points to Mention

  • Computational efficiency: two-stage reduces the number of expensive scoring operations by filtering with a cheap model first.
  • Latency: two-stage can have higher latency due to sequential stages, but each stage is faster; single-pass latency is predictable but may be high for large sets.
  • Accuracy: two-stage can achieve higher accuracy by using a powerful reranker on a small set, but risks missing relevant items if the first stage has low recall; single-pass may compromise accuracy for speed.
  • Scalability: two-stage scales better to large candidate sets because the expensive model only sees a subset.
  • Complexity: two-stage requires training and maintaining two models and a retrieval mechanism, while single-pass is simpler.
  • Use cases: two-stage is common in search and recommendation systems with large item catalogs; single-pass is suitable for smaller sets or when latency is critical.

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

Q9

What are the tradeoffs between a two-stage and a single-stage ranking architecture?

System DesignTechnical Trade-offs
Author's notes

Overlapped a lot with the previous question.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining what each architecture entails and then systematically compare them across key dimensions like latency, complexity, and business impact. Use a concrete example (e.g., search or recommendation) to illustrate the tradeoffs and conclude with when to choose each based on constraints.

Pro tip: Tie the tradeoffs to Credit Karma's context: a two-stage system can handle large-scale candidate generation efficiently, but a single-stage model may be simpler and sufficient if the candidate set is small or latency budgets are tight. Always mention monitoring and iteration costs, as these often tip the decision in production.

1. Define the architectures

Briefly explain that a single-stage ranking uses one model to score all items, while a two-stage system first retrieves a candidate set (e.g., via a lightweight model) and then ranks them with a more complex model.

2. Compare on latency and scalability

Discuss how two-stage can reduce latency by pruning candidates early, but adds overhead from the retrieval stage; single-stage may be faster for small item sets but struggles with millions of items.

3. Compare on model complexity and accuracy

Explain that two-stage allows a simple retrieval model and a sophisticated ranker, potentially improving accuracy, while single-stage must balance complexity and speed in one model, often leading to compromises.

4. Consider operational and business factors

Mention maintenance, debugging, and iteration speed: two-stage is harder to maintain and tune but offers flexibility; single-stage is simpler but may not scale. Also relate to business metrics like conversion or user engagement.

5. Conclude with recommendations

Summarize when to use each: two-stage for large-scale, low-latency systems with diverse candidates; single-stage for smaller scale, simpler systems, or when latency is critical and candidate set is manageable.

Key Points to Mention

  • Latency vs. accuracy tradeoff: two-stage can achieve lower latency by reducing the number of items scored by the complex model, but may sacrifice recall if retrieval is poor.
  • Scalability: two-stage handles large item catalogs better by using approximate nearest neighbor or simple models for retrieval.
  • Model complexity: single-stage requires one model to handle all items, which can be limiting; two-stage allows specialized models for retrieval and ranking.
  • Maintenance and iteration: two-stage systems are more complex to develop, debug, and update, requiring coordination between stages.
  • Business impact: consider how each affects key metrics like click-through rate, conversion, and user satisfaction, and the cost of errors.
  • Cold start and diversity: two-stage can incorporate diversity and exploration in retrieval, while single-stage may struggle to balance these with ranking.

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

Q10

Under what conditions would you choose a two-stage architecture over a single-stage one?

System DesignTechnical Trade-offsAdaptability & Ambiguity
Author's notes

Corpus size, latency constraints, feature cost at scoring time.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying that the choice depends on the problem's complexity, data characteristics, and business constraints. Then, walk through the key trade-offs between two-stage and single-stage architectures, using concrete examples from your experience. Finally, emphasize that the decision should be driven by empirical evaluation and iterative refinement.

Pro tip: Frame your answer around the specific context of the company (e.g., Credit Karma's need for interpretable and scalable models) and mention how you would measure success (e.g., latency, accuracy, maintainability) to show business acumen.

1. Clarify the Problem and Constraints

Ask about the specific use case, data volume, latency requirements, and interpretability needs to ground your answer in the actual scenario.

2. Define Single-Stage and Two-Stage Architectures

Briefly explain what each architecture entails (e.g., single-stage: end-to-end model; two-stage: separate retrieval/ranking or detection/classification) to ensure shared understanding.

3. Compare Trade-offs

Discuss key dimensions such as accuracy, latency, scalability, maintainability, and data requirements, highlighting when two-stage might outperform single-stage.

4. Provide Concrete Examples

Illustrate with real-world scenarios (e.g., recommendation systems, fraud detection) where two-stage is preferred, and explain why.

5. Conclude with Decision Criteria

Summarize the conditions under which you would choose two-stage, and mention the importance of experimentation and monitoring to validate the choice.

Key Points to Mention

  • Problem complexity and modularity: two-stage allows decomposing a complex task into simpler sub-problems.
  • Data availability and quality: two-stage can handle heterogeneous data or when labeled data is scarce for end-to-end training.
  • Latency and throughput: two-stage may introduce additional latency but can be optimized with caching or parallel processing.
  • Interpretability and debugging: two-stage provides intermediate outputs that are easier to interpret and debug.
  • Scalability and maintenance: two-stage enables independent scaling and updating of components.
  • Empirical evaluation: the choice should be validated through A/B testing and metrics aligned with business goals.

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