← Upstart Interview Insights

Upstart·Data Scientist·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say
May 2026

Summary

System design round at Upstart for a data scientist role. The whole thing was one big open-ended question about ads serving, and they really wanted you to go deep on every layer of the stack. Felt like a senior-level bar even if the title didn't say so.

Questions Asked (1)

Q1

Walk through the end-to-end design of an ads serving system for a social media app, covering data collection, feature engineering, model selection, real-time ranking, feedback loops, and A/B evaluation. Also discuss offline vs online metrics and how you'd handle cold-start users.

System DesignA/B Testing & ExperimentationTechnical Trade-offs
Author's notes

This is basically a dissertation prompt.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as a real-time recommendation system with a feedback loop, then walk through each component (data, features, models, serving, evaluation) in a logical order. Emphasize trade-offs between latency, accuracy, and scalability, and tie your choices back to business metrics like CTR and revenue. Conclude by discussing how you'd iterate using A/B tests and handle cold-start users.

Pro tip: Show that you understand the end-to-end system isn't just about the model—it's about the data pipeline, serving infrastructure, and experimentation culture. Mention concrete examples of how you'd monitor and mitigate issues like feedback loops or position bias.

1. Data Collection & Feature Engineering

Describe the data sources (user interactions, ad content, context) and how you'd process them into features for training and serving. Highlight the need for consistent feature computation offline and online.

2. Model Selection & Training

Discuss candidate models (e.g., logistic regression, GBDT, deep learning) and why you'd choose one based on latency, interpretability, and performance. Explain how you'd train and validate the model.

3. Real-Time Ranking & Serving

Outline the serving architecture: candidate generation, feature fetching, model inference, and ranking. Address latency constraints and fallback strategies.

4. Feedback Loops & Online Learning

Explain how you'd collect feedback (clicks, conversions) and use it to update models, either in batch or online. Discuss challenges like feedback loops and how to mitigate them.

5. A/B Testing & Evaluation

Describe how you'd design A/B tests to measure the impact of model changes, including offline metrics (AUC, log loss) and online metrics (CTR, revenue). Discuss cold-start handling and guardrail metrics.

Key Points to Mention

  • Feature store for consistency between offline and online features
  • Trade-offs between model complexity and latency in real-time serving
  • Handling position bias and feedback loops in training data
  • Offline metrics (e.g., AUC, NDCG) vs online metrics (CTR, CVR, revenue)
  • Cold-start strategies: content-based features, exploration (epsilon-greedy, Thompson sampling), and using side information
  • A/B testing best practices: randomization unit, sample size, guardrail metrics, and sequential testing

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