Machine Learning interviews

Work in the tool while you explain your decisions.

Choose a role and level. Each interview has a new task, follow-up questions and a report that checks your work and your explanation.

Take the assessment Choose an interview

Speak or type. Sign in to save your results across devices.

Timed assessment

12 practical questions
50 minutes
Mid level

Questions are drawn from a bank of 36. Your report includes a skill breakdown and lessons to practice next.

Start assessment

Practice interviews

Junior tasks cover the fundamentals. Mid and senior tasks add debugging, changing requirements and review decisions.

Position

10 interviews

Data scientist
Junior
35 min

Juniper Subscriptions: a reproducible customer holdout

Juniper Subscriptions is a fictional business using the supplied original measurements. Split row_ids into train_ids and test_ids with test_size=0.3, random_state=37 and stratify=y. Keep the original row identities so you can audit membership.

Start interview

Skills and rubricDisjoint stratified holdouts

Define the unit being split and why one customer must not leak across partitions.
Split original row identities before preprocessing. A fixed seed makes this audit reproducible; stratification preserves class representation.
Check intersection, union and class counts against the source identities.

Machine learning engineer
Junior
35 min

Tern Dispatch: a demand baseline before tuning

Tern Dispatch is a fictional business using the supplied original measurements. Fit baseline as DummyRegressor(strategy="median") on X_train and yr_train. Store validation_mae from its predictions for X_test against yr_test. Keep evaluation targets out of fit.

Start interview

Skills and rubricTraining-only median baseline

State what the demand forecast changes for the planner before comparing models.
A baseline establishes what a model must improve. The training median is the constant prediction that minimizes training absolute error.
Recompute MAE from held-out predictions and inspect the fitted constant.

Data scientist
Junior
35 min

Clover Retail: a fitted dispatch regression

Clover Retail is a fictional business using the supplied original measurements. Fit model as LinearRegression on X_train and yr_train. Use the fitted model to predict X_test. The checker reads the estimator coefficients and unseen-row predictions.

Start interview

Skills and rubricFitted linear prediction

Identify the forecast target and when each dispatch feature becomes available.
Fit coefficients on training measurements. Predict evaluation rows with that same estimator and compare it with an appropriate baseline.
Inspect held-out residuals and check that the intercept is fitted rather than supplied.

Machine learning engineer
Mid
35 min

Mossline Media: a churn pipeline with one fitted scaler

Mossline Media is a fictional business using the supplied original measurements. Fit model as StandardScaler followed by LogisticRegression(C=1, max_iter=1000) in a pipeline on X_train and y_train. Keep preprocessing inside the pipeline. The checker reads fitted scaler statistics and unseen probabilities.

Start interview

Skills and rubricTraining-only scaling and classification

Fit preprocessing and classification together on training data. Evaluation observations receive transform and predict, without changing fitted statistics.
Where is the scaler refitted when this pipeline enters cross-validation?
Read the scaler training means and validate probabilities on rows never used in fit.

Data scientist
Mid
35 min

Alder Planning: a regularized demand model

Alder Planning is a fictional business using the supplied original measurements. Fit model as StandardScaler followed by Ridge(alpha=4) on X_train and yr_train. The checker reads the fitted ridge coefficients, training means and evaluation predictions.

Start interview

Skills and rubricScaled ridge regularization

Scale before comparing coefficient penalties. Alpha controls shrinkage and must be selected without using final test outcomes.
Why does changing feature units matter for an unscaled ridge penalty?
Compare validation errors across training folds and keep the final test for one evaluation.

Machine learning engineer
Mid
35 min

Brookline Payments: a bounded forest review

Brookline Payments is a fictional business using the supplied original measurements. Fit model as RandomForestClassifier with n_estimators=40, max_depth=3 and random_state=23 on X_train and y_train. Read unseen probabilities from the fitted forest.

Start interview

Skills and rubricControlled forest fitting

Limit complexity for a reproducible comparison. More trees and deeper trees change different parts of a forest and don't replace honest evaluation.
Why can adding trees stabilize predictions without fixing leakage?
Compare held-out segment errors and baseline results before accepting the forest.

Data scientist
Mid
35 min

Quarry Forecasting: a search with an untouched final test

Quarry Forecasting is a fictional business using the supplied original measurements. Fit search as GridSearchCV of StandardScaler then LogisticRegression(max_iter=1000), searching logisticregression__C over [.05,.5,5]. Use StratifiedKFold(3, shuffle=True, random_state=29), scoring="neg_log_loss" and training data only.

Start interview

Skills and rubricPipeline search within training folds

Cross-validation refits the entire pipeline in each training fold. Search scores select a configuration; the untouched final test estimates its later performance.
How would repeated account IDs change the stratified folding plan?
Inspect the scoring objective and fold memberships, then evaluate the selected model once.

Machine learning engineer
Senior
35 min

Lark Review: a fraud queue operating threshold

Lark Review is a fictional business using the supplied original measurements. Set predicted = probabilities >= threshold. Store report as a NumPy array containing precision, recall and F1 in that order, using actual as the true labels and zero_division=0 for each metric.

Start interview

Skills and rubricPrecision, recall and F1 at a chosen threshold

Review capacity falls while fraud prevalence rises. Which threshold evidence must be recomputed?
Reconcile predicted positives, actual positives and true positives independently.
Define an acceptance owner, a monitoring limit and a reversible response when evaluation evidence changes.

Data scientist
Senior
35 min

Cove Services: calibrated probabilities for prioritization

Cove Services is a fictional business using the supplied original measurements. Fit model as CalibratedClassifierCV of a StandardScaler/LogisticRegression(max_iter=1000) pipeline, method="sigmoid", cv=3, on X_train and y_train. Inspect evaluation probabilities from this fitted calibrator.

Start interview

Skills and rubricCalibration with disjoint folds

The class prevalence changes after launch. Which reliability evidence might change even if ranking holds?
Inspect reliability bins with counts and proper scoring rules on mature held-out labels.
Define an acceptance owner, a monitoring limit and a reversible response when evaluation evidence changes.

Machine learning engineer
Senior
35 min

Reed Logistics: held-out feature importance

Reed Logistics is a fictional business using the supplied original measurements. Fit model as RandomForestClassifier(n_estimators=40, max_depth=3, random_state=23) on X_train and y_train. Store importance from permutation_importance on X_test and y_test, scoring="accuracy", n_repeats=4, random_state=31.

Start interview

Skills and rubricPermutation importance on unseen data

A manager calls the leading feature a cause of churn. What evidence is missing?
Repeat the permutation protocol on held-out rows and retain its spread and segment context.
Define an acceptance owner, a monitoring limit and a reversible response when evaluation evidence changes.

Choose your character

Machine Learning interviews

Timed assessment

Practice interviews

Juniper Subscriptions: a reproducible customer holdout

Tern Dispatch: a demand baseline before tuning

Clover Retail: a fitted dispatch regression

Mossline Media: a churn pipeline with one fitted scaler

Alder Planning: a regularized demand model

Brookline Payments: a bounded forest review

Quarry Forecasting: a search with an untouched final test

Lark Review: a fraud queue operating threshold

Cove Services: calibrated probabilities for prioritization

Reed Logistics: held-out feature importance

Interview Hub access

Your Work Gym account