Coinbase·Machine Learning Engineer·Technical Phone Screen·Senior

Senior

Jun 2026Remote

Summary

Coinbase ML Engineer interview where they dropped a messy Jupyter notebook in front of me and said go. One session, real data, binary classification, and a clock ticking in the background. Not a vibe check, they actually wanted working code.

Questions Asked (1)

Given a messy real-world tabular dataset in a live notebook, build a working end-to-end binary classification baseline: clean the data, split without leakage, train a model that handles mixed feature types, evaluate with an appropriate metric, and explain your choices.

Technical Trade-offsData ModelingAlgorithms & Data Structures

Author's notes

The part that tripped me up first was the identifier columns.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by profiling the data to identify quality issues and feature types, then systematically clean and preprocess using a pipeline that handles mixed types. Split the data before any preprocessing to avoid leakage, train a simple baseline model, and evaluate using appropriate metrics while explaining each choice in the context of the business problem.

Pro tip: Always set up a reproducible pipeline with fixed random seeds and document every decision—this demonstrates production readiness and makes your baseline easily iterable. Also, consider the cost of false positives vs. false negatives early to guide metric selection.

1. Data Profiling and Cleaning

Inspect the dataset for missing values, outliers, duplicates, and inconsistencies. Decide on imputation or removal strategies and document your reasoning.

2. Leakage-Safe Splitting

Split the data into train/validation/test sets before any preprocessing. Use stratified splitting if classes are imbalanced to maintain distribution.

3. Preprocessing Pipeline

Build a pipeline that applies appropriate transformations to numerical (e.g., scaling) and categorical (e.g., one-hot encoding) features. Fit the pipeline only on training data.

4. Model Training and Evaluation

Train a simple baseline model (e.g., logistic regression or random forest) that handles mixed types. Evaluate using metrics like AUC-ROC, precision-recall, or F1, depending on class balance and business costs.

5. Explain Choices and Iterate

Clearly articulate why you chose each step, from cleaning to metric. Suggest next steps for improvement, such as feature engineering or hyperparameter tuning.

Key Points to Mention

Data leakage prevention: split before preprocessing and fit transformations only on training data.
Handling mixed feature types: use ColumnTransformer to apply different preprocessing to numerical and categorical columns.
Choice of evaluation metric: consider class imbalance and business impact (e.g., precision-recall for rare events).
Baseline model selection: start simple (logistic regression) for interpretability, then consider tree-based models for mixed data.
Reproducibility: set random seeds and document all preprocessing steps.
Iterative improvement: baseline is a starting point; plan for feature engineering and model tuning.

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