Machine Learning Engineer Interview Questions

What should you focus on for a Machine Learning Engineer interview?

What makes these interviews different?

ML fundamentals, evaluation, and deep learning stand out. These topics appeared more often in these interviews than across interviews for all roles. Computer vision and model deployment also appeared more often than average. Most questions were junior or mid-level, with senior-level questions too.

Start by checking both concepts and decisions: can you explain how a model learns, how you would test it, and how you would run it in production?

Can you explain model errors and test results?

Overfitting and underfitting was the most asked question, appearing in 6 of 69 interviews. Questions also covered learning approaches and preprocessing leakage. Check whether you can answer these without stopping at a definition:

• How do you detect and prevent overfitting and underfitting?
• When would you use supervised, unsupervised, or reinforcement learning?
• How do you prevent preprocessing leakage in scikit-learn cross-validation, including time-series splits?
• How would you choose a metric, validation split, and decision threshold?

Can you explain neural networks and vision tasks?

The questions included CNNs, Transformers, multilayer perceptrons, convolution, and backpropagation. Check each topic separately: can you explain CNN and Transformer architectures, how convolution works, and how backpropagation works?

Other questions covered losses, optimizers, and regularization. Can you explain your choices for model training? If labeled data is scarce, can you describe how you would use data augmentation and transfer learning?

For computer vision, can you explain how you would prepare images for detection or segmentation, or track objects in video?

Can you explain Python behavior and write code?

Tasks that required writing code appeared in 6 of 69 interviews. The recorded tasks give you concrete checks beyond knowing Python syntax:

• Can you count character frequencies in a user-provided name?
• Can you check whether a string is a palindrome?
• For an LRU cache, what changes when a key already exists? When capacity is exceeded, which node do you remove, and how do you update the linked list and hash map?

Also check NumPy: can you explain views versus copies and test whether arrays share memory?

Can you explain production and project decisions?

Deployment questions covered serving, monitoring, retraining, and reliability. Can you explain how you would keep training and serving consistent, speed up inference, and handle more traffic?

For project discussions, can you explain a model choice in terms of business trade-offs? Can you discuss responsible AI and decisions made with stakeholders?

What Machine Learning Engineer interviews cover

The 20 most asked Machine Learning Engineer questions

Counted by interview, not by mention: a question asked twice in one conversation counts once.

1. What are overfitting and underfitting, and how do you detect and prevent them?

   Asked in 6 of 69 interviews · last asked Oct 19 · Concept · Mid-level

   Show answer

   Overfitting means learning training details or noise that do not generalize; underfitting means missing important patterns. Overfitting gives low training error but higher validation error; underfitting gives high error on both. Reduce overfitting with regularization, early stopping, simpler models, or more representative data. Reduce underfitting with better features, more model capacity, or less regularization.

   • Compare training and validation errors using a split that reflects how the model will be used.
   • Learning curves help reveal overfitting, underfitting, and whether more training may help.
   • Use validation data or cross-validation to choose model settings; keep test data for final evaluation.
   • Prevent data leakage, which can make validation results look better than they really are.

   What the interviewer checks: The interviewer checks whether you can diagnose poor generalization and choose a suitable fix.

   Other ways it was asked

   • “Describe overfitting, how you detect it, and how you prevent it.”
   • “Can you explain the difference between underfitting and overfitting?”
   • “What are underfitting and overfitting, and how do they differ?”

2. How do supervised, unsupervised, and reinforcement learning differ, and when would you use each?

   Asked in 4 of 69 interviews · last asked Oct 19 · Concept · Junior

   Show answer

   Supervised learning uses labeled examples to predict outputs, such as spam labels or house prices. Unsupervised learning finds patterns without target labels, such as customer groups. Reinforcement learning learns how to act from rewards, aiming to maximize total reward over time; I use it for sequential decisions, such as robot control.

   • Choose supervised learning when you have labeled data and a clear prediction target.
   • Choose unsupervised learning to explore structure when target labels are unavailable.
   • Choose reinforcement learning when actions affect future outcomes and rewards provide feedback.
   • Reinforcement learning may require exploration, which can be costly or unsafe in real systems.

   What the interviewer checks: The interviewer checks whether you can match each learning method to its feedback type and suitable tasks.

   Other ways it was asked

   • “What are the differences between supervised, unsupervised, and reinforcement learning?”
   • “Could you explain the differences between supervised, unsupervised, and reinforcement learning?”
   • “Describe the key differences between supervised, unsupervised, and reinforcement learning, and when each is appropriate.”

3. Can you explain the CNN architecture?

   Asked in 3 of 69 interviews · last asked Sep 23 · Concept · Junior

   Show answer

   A convolutional neural network (CNN) learns patterns in images using filters that slide across the input. Convolution layers build feature maps, and activation functions help learn complex patterns. Pooling or strided convolutions can reduce spatial size. Later layers combine features, and an output layer makes the prediction.

   • Filters share weights across image locations, reducing parameters compared with fully connected layers.
   • Early layers often learn edges; deeper layers combine them into more complex patterns.
   • Pooling is optional; strided convolutions can also reduce spatial size.
   • The output depends on the task, such as image classification or pixel-level segmentation.

   What the interviewer checks: The interviewer checks whether you understand the main CNN layers, weight sharing, and how image features become predictions.

   Other ways it was asked

   • “Can you explain CNN architecture?”
   • “How does a CNN model work?”

4. How do you prevent preprocessing leakage in scikit-learn cross-validation, including time-series splits?

   Asked in 3 of 69 interviews · last asked Jun 3 · Tooling · Mid-level

   Show answer

   Put all learned preprocessing and the model in one Pipeline, then cross-validate that pipeline. Use ColumnTransformer for different column types; each fold fits preprocessing only on its training data. For time series, sort by time and use TimeSeriesSplit, with a gap when needed. Features must use only information available at prediction time.

   • Never fit imputers, scalers, encoders, or feature selectors on the full dataset before cross-validation.
   • Validation data uses the preprocessing fitted on that fold’s training data.
   • Choose the time-series gap to prevent training labels from overlapping the validation period.
   • A pipeline cannot fix features that already contain future information or leaked targets.

   PythonCopy

   from sklearn.compose import ColumnTransformer
   from sklearn.impute import SimpleImputer
   from sklearn.preprocessing import StandardScaler, OneHotEncoder
   from sklearn.pipeline import make_pipeline
   from sklearn.linear_model import Ridge
   from sklearn.model_selection import TimeSeriesSplit, cross_val_score
   num = make_pipeline(SimpleImputer(strategy="median"), StandardScaler())
   prep = ColumnTransformer([("num", num, numeric_cols), ("cat", OneHotEncoder(handle_unknown="ignore"), categorical_cols)])
   model = make_pipeline(prep, Ridge(solver="lsqr"))
   scores = cross_val_score(model, X, y, cv=TimeSeriesSplit(n_splits=5, gap=gap))

   What the interviewer checks: The interviewer checks whether you keep learned preprocessing inside each fold and respect what information is available over time.

   Other ways it was asked

   • “In scikit-learn, how do you correctly run cross-validation when preprocessing steps such as scaling and one-hot encoding are involved, to avoid data leakage?”
   • “In scikit-learn, how do you avoid data leakage when doing preprocessing and cross-validation, especially with scaling and imputation?”
   • “In scikit-learn, how do you prevent target leakage when doing time-series cross-validation with feature scaling and encoding, and what APIs do you use to enforce that?”

5. What is the Transformer architecture?

   Asked in 3 of 69 interviews · last asked Aug 27 · Concept · Junior

   Show answer

   A Transformer is a neural network architecture that uses attention to learn relationships between tokens in a sequence. Each layer combines attention with a feed-forward network, residual connections, and normalization. Position information tells the model about token order. The original design has an encoder and a decoder, but encoder-only and decoder-only models are also common.

   • Self-attention uses queries, keys, and values to build token representations from relevant context.
   • Multiple attention heads can learn different relationships between tokens.
   • Causal masking prevents a decoder from attending to future tokens.
   • Standard full attention has quadratic cost in sequence length; training is parallelizable, but autoregressive generation is sequential.

   What the interviewer checks: The interviewer checks whether you understand attention, position information, the main building blocks, and common Transformer variants.

   Other ways it was asked

   • “What is transformer architecture, and how does it work?”
   • “What are transformer architectures, and how do they work?”

6. How do NumPy views differ from copies, and how can you check whether arrays share memory?

   Asked in 2 of 69 interviews · last asked Jun 3 · Concept · Mid-level

   Show answer

   A NumPy view shares data with another array, while a copy has its own data buffer. Changing a view can change the original, so use `.copy()` when you need independent data. Use `np.shares_memory(a, b)` to check whether two arrays overlap in memory.

   • Basic slicing returns views; integer-array and boolean indexing return copies.
   • `np.shares_memory` checks actual overlap, but can be expensive for complex layouts.
   • `np.may_share_memory` is conservative: True means possible overlap, not certain overlap.
   • The `.base` attribute alone does not prove that two arrays overlap in memory.

   PythonCopy

   import numpy as np
   
   a = np.arange(5)
   view = a[1:4]
   copy = view.copy()
   print(np.shares_memory(a, view))  # True
   print(np.shares_memory(a, copy))  # False
   view[0] = 99
   print(a)  # [ 0 99  2  3  4]

   What the interviewer checks: The interviewer is checking whether you understand shared data and can prevent accidental changes to the original array.

   Other ways it was asked

   • “In NumPy, how does a view differ from a copy, and how can you tell which one you got after slicing?”
   • “In NumPy, what's the difference between a view and a copy, and how do you prevent accidental data mutation when slicing?”
   • “How would you check in code whether a sliced NumPy array shares memory with the original array?”

7. What is a multilayer perceptron, and how does it work?

   Asked in 2 of 69 interviews · last asked Oct 15 · Concept · Junior

   Show answer

   A multilayer perceptron (MLP) is a feedforward neural network with an input layer, one or more hidden layers, and an output layer. Each hidden layer applies learned weights, biases, and a nonlinear activation. During training, backpropagation computes gradients, and an optimizer updates the parameters to reduce loss. MLPs can learn complex patterns for classification or regression.

   • Layers are typically fully connected: each neuron receives inputs from every neuron in the previous layer.
   • Nonlinear activations let the network learn more than a linear mapping.
   • Choose the output activation and loss for the task, such as softmax and cross-entropy for multiclass classification.
   • Scale numeric inputs and check validation performance to detect overfitting.

   What the interviewer checks: The interviewer checks whether you understand the network structure, forward computation, and how training updates its parameters.

   Other ways it was asked

   • “What is a multilayer perceptron, and when would you use one?”
   • “How does a multilayer perceptron work, and how do you solve a problem using one?”

8. How does convolution work in a neural network's convolutional layer?

   Asked in 2 of 69 interviews · last asked Sep 23 · Concept · Mid-level

   Show answer

   A convolutional layer slides a learned filter across the input. At each position, it multiplies input values by filter weights, sums the results, and usually adds a bias. This forms a feature map. Reusing the same weights across positions helps detect patterns, such as edges, wherever they appear.

   • Each filter spans the input channels it uses and produces one output feature map.
   • Training learns the filter weights; sharing weights reduces parameter count.
   • Stride controls movement between positions; padding controls how borders are handled and affects output size.
   • Most libraries use cross-correlation: they do not flip the filter as mathematical convolution does.

   What the interviewer checks: The interviewer checks whether you understand local filtering, shared weights, and how feature maps are produced.

   Other ways it was asked

   • “What is convolution, and how is it used in neural networks?”
   • “What is convolution, and how does it work in a neural network?”
   • “Can you explain how a convolutional layer works in deep learning?”

9. How would you handle scarce labeled data, including using data augmentation and transfer learning?

   Asked in 2 of 69 interviews · last asked Oct 17 · Scenario · Mid-level

   Show answer

   I would use a model trained on related data and fine-tune it on my labeled examples. I would also create extra training examples with changes that preserve the correct label. If needed, I would label more useful examples or use unlabeled data. I would check results on a separate, representative validation set.

   • Transfer learning reuses learned features; freeze layers or fine-tune carefully to reduce overfitting.
   • Augmentation must preserve labels and match realistic inputs; some changes can alter the meaning.
   • Split data before augmentation, and augment only training data to prevent leakage.
   • Use active learning to choose useful examples for labeling; check pseudo-labels for errors.

   What the interviewer checks: The interviewer is checking whether you can learn from limited labels while controlling overfitting and data leakage.

   Other ways it was asked

   • “Regarding supervised learning, how do you handle situations where labeled data is scarce?”
   • “How do data augmentation and transfer learning help when labeled data is scarce in supervised learning?”
   • “What would you do if labeled data were scarce for a machine learning problem?”

10. Describe some common machine learning architectures and when you would use them.

    Asked in 2 of 69 interviews · last asked Oct 19 · Concept · Mid-level

    Show answer

    I use tree ensembles or feedforward neural networks for table data. Convolutional neural networks (CNNs) suit images because they learn local patterns. Recurrent neural networks (RNNs), including LSTMs, process sequences step by step, while transformers use attention for text and other data. I choose based on the task, data size, speed, and compute limits.

    • Tree ensembles are strong baselines for table data, but do not naturally model image or sequence structure.
    • CNNs capture spatial patterns well, but may struggle with relationships between distant image regions.
    • RNNs handle ordered data, but sequential processing limits training speed and long-range memory.
    • Transformers capture relationships across positions, but standard attention becomes expensive for long inputs.

    What the interviewer checks: The interviewer checks whether you can match model structures to tasks and explain their main trade-offs.

    Other ways it was asked

    • “Describe common machine learning architectures, their typical use cases, and briefly note any limitations.”

11. Can you explain the backpropagation method?

    Asked in 2 of 69 interviews · last asked Sep 23 · Concept · Junior

    Show answer

    Backpropagation computes how much each model parameter affects the loss. After a forward pass, it works backward through the network, using the chain rule to calculate gradients. An optimizer then uses these gradients to update the parameters and try to reduce the loss.

    • The forward pass computes predictions and loss, saving values needed for gradients.
    • The backward pass applies the chain rule from the loss toward earlier layers.
    • Backpropagation computes gradients; the optimizer updates weights and biases.

    What the interviewer checks: The interviewer is checking whether you understand the chain rule and can separate gradient calculation from parameter updates.

    Other ways it was asked

    • “Explain backpropagation.”

12. How would you design validation splits to prevent leakage from time ordering, repeated users, or related observations?

    Asked in 2 of 69 interviews · last asked Aug 10 · Scenario · Senior

    Show answer

    I design splits to match how the model will be used. For future predictions, I train on earlier data and validate on later data, using only information available at prediction time. For unseen users or independent groups, I keep each user or group in one split. I fit preprocessing and feature selection only on training data.

    • Use rolling time splits to test performance across several future periods.
    • Keep duplicates and related observations together when they must be independent across splits.
    • Past interactions from validation users are valid when predicting future interactions for known users.
    • Add a time gap when overlapping label windows could leak future information.

    What the interviewer checks: The interviewer checks whether you can match validation to deployment and prevent information from crossing split boundaries.

    Other ways it was asked

    • “How would you structure validation splits to prevent data leakage or look-ahead bias?”
    • “How do you design validation when there's a risk of leakage, such as with time-series data, multiple rows per user, or repeated interactions in a recommender system?”

13. Can you explain a confusion matrix?

    Asked in 2 of 69 interviews · last asked Aug 27 · Concept · Junior

    Show answer

    A confusion matrix shows how a classifier’s predictions compare with the true labels. For binary classification, it counts true positives, true negatives, false positives, and false negatives. It helps us see which mistakes the model makes, not just how often it is correct.

    • True positives and true negatives are correct predictions for the positive and negative classes.
    • False positives are false alarms; false negatives are missed positive cases.
    • Check axis labels: rows and columns can represent actual or predicted classes.
    • For multiple classes, each cell counts cases with a specific actual class and predicted class.

    What the interviewer checks: The interviewer is checking whether you can explain classification errors and read a confusion matrix correctly.

    Other ways it was asked

    • “What do you understand by a confusion matrix?”
    • “What are the components of a confusion matrix?”
    • “Explain the confusion matrix.”

14. What is the difference between training data and testing data?

    Asked in 2 of 69 interviews · last asked Aug 21 · Concept · Junior

    Show answer

    Training data is used to teach a model by fitting its parameters. Test data is kept separate and used to measure how well the trained model works on unseen examples. I use validation data, not test data, to choose models and tune settings.

    • Training data guides learning; test data measures performance on unseen examples.
    • Keep test data out of training and tuning to avoid misleading results.
    • Split data to match real use, keeping related examples or future information from leaking across splits.

    What the interviewer checks: The interviewer is checking whether you understand learning, evaluation, and how to avoid data leakage.

    Other ways it was asked

    • “What is the difference between training data and test data?”

15. What is an F1 score?

    Asked in 2 of 69 interviews · last asked Sep 29 · Concept · Junior

    Show answer

    The F1 score is the harmonic mean of precision and recall: F1 = 2 × precision × recall / (precision + recall). It balances false positives and false negatives, with higher scores being better. It is useful for imbalanced classes, but it ignores true negatives.

    • Precision measures how many predicted positives are correct; recall measures how many actual positives are found.
    • F1 ranges from 0 to 1 and is high only when both precision and recall are high.
    • For multiple classes, specify the averaging method, such as macro, micro, or weighted.
    • If there are no actual or predicted positives, F1 is undefined; libraries use configurable handling.

    What the interviewer checks: The interviewer is checking whether you understand how F1 combines precision and recall and when it is useful.

    Other ways it was asked

    • “What is the F1 score?”

16. How would you connect to an RTSP camera stream and extract frames for computer vision processing?

    Asked in 2 of 69 interviews · last asked Sep 22 · Tooling · Mid-level

    Show answer

    I would open the camera’s RTSP URL with OpenCV using an FFmpeg backend that supports RTSP. I would read decoded frames, preprocess them, and pass them to the vision model. For continuous use, I would handle connection failures, reconnect with a delay, and drop old frames if processing falls behind.

    • Use the camera’s correct RTSP path and credentials; keep passwords out of source code and logs.
    • Check that the stream opens and each frame read succeeds.
    • OpenCV normally returns BGR frames; convert to RGB if the model requires it.
    • Limit queued frames to reduce delay, and release the connection when finished.

    PythonCopy

    import cv2
    cap = cv2.VideoCapture("rtsp://camera-host/stream", cv2.CAP_FFMPEG)
    try:
        if not cap.isOpened():
            raise RuntimeError("Cannot open RTSP stream")
        while True:
            ok, frame = cap.read()
            if not ok: break
            gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
    finally: cap.release()

    What the interviewer checks: The interviewer checks whether you can connect video input to model processing while handling failures and delay.

    Other ways it was asked

    • “How would you extract frames from an RTSP stream for face recognition?”
    • “How do you process an RTSP video stream?”
    • “How would you connect a computer vision application to an RTSP stream from a camera?”

17. How would you evaluate face recognition accuracy, and what data, tasks, and metrics would you use?

    Asked in 2 of 69 interviews · last asked Sep 22 · Scenario · Mid-level

    Show answer

    I would test face verification and identification separately on held-out, representative data. Verification asks whether two faces match; identification searches a gallery. I would report error rates, not just overall accuracy. I would also compare results across demographic groups, lighting, pose, and image quality.

    • Use identity-disjoint training, validation, and test sets; include matching pairs, nonmatching pairs, and realistic galleries.
    • For verification, report ROC curves and false-reject rates at chosen false-accept rates; tune thresholds on validation data.
    • For identification, report rank-one and top-k accuracy; when unknown people appear, measure false identifications and missed matches.
    • Use established benchmarks, such as LFW or IJB-C, but also test data that reflects the intended use.

    What the interviewer checks: The interviewer checks whether you can match the evaluation data and metrics to the task and real-world error costs.

    Other ways it was asked

    • “How should you measure and interpret the accuracy of a face-recognition system such as DeepFace?”
    • “What data or evaluation task did you use to measure a face recognition model's accuracy?”

18. How should a deployed model adapt or gracefully degrade when incoming data or user behavior changes?

    Asked in 2 of 69 interviews · last asked Sep 14 · Scenario · Senior

    Show answer

    A deployed model should monitor changes in inputs and outcomes, then adapt only after validation. I would check performance with recent labels and test retrained models before a gradual rollout. If quality drops or inputs are unsupported, I would use a tested fallback, limit predictions, or route cases to human review.

    • Data drift is a warning, not proof that model accuracy has dropped.
    • Track performance across user groups; delayed labels can hide failures.
    • Retrain on representative recent data and validate before deployment.
    • Use safe rules, a simpler model, or human review when predictions are unreliable.

    What the interviewer checks: The interviewer checks whether you can detect change, update safely, and protect users when predictions become unreliable.

    Other ways it was asked

    • “If data patterns or human behavior shift after deployment, how would you adapt the system or make it degrade gracefully?”
    • “What can we do when the model receives different data after deployment?”

19. How do you monitor production models and set alerts for drift, performance degradation, retraining, and rollbacks?

    Asked in 2 of 69 interviews · last asked May 24 · Scenario · Senior

    Show answer

    I monitor input quality, drift, prediction quality, calibration, latency, and errors by model version and user group. I set alerts against validation baselines and service targets, using sustained breaches and enough samples. Drift triggers investigation; confirmed quality loss can trigger retraining. I roll back to a known-good version when a release harms quality or breaks service targets.

    • Compare feature distributions and missing-value rates with training data and stable production periods.
    • With delayed labels, use early signals, then measure task metrics and calibration when labels arrive.
    • Set task-specific thresholds for quality, calibration, p95 latency, and error rates; there are no universal cutoffs.
    • Validate retrained models on recent data, then test on limited traffic before replacing the current model.

    What the interviewer checks: The interviewer checks whether you can detect meaningful problems and choose safe actions without treating every drift alert as failure.

    Other ways it was asked

    • “Once a model is in production, what monitoring do you set up for data drift, performance, calibration, and latency, and what thresholds or triggers cause you to retrain or roll back?”
    • “How would you implement model monitoring to detect data drift and performance degradation in production, and what metrics would you alert on?”

20. Describe a project you built with Python and machine learning frameworks, including challenges and outcomes.

    Asked in 2 of 69 interviews · last asked Aug 26 · Behavioral · Mid-level

    Show answer

    Hypothetical example: I built a support ticket classifier in Python with scikit-learn to suggest the right team. I used TF-IDF and logistic regression, with class weights for rare teams. I removed duplicates and tested on newer tickets using a time-based split. It beat a majority-class baseline on macro F1, but rare teams still needed human review.

    • Explain the problem, your own contribution, and why you chose the framework.
    • Describe a real challenge and how you handled it.
    • Compare results with a baseline using a metric that fits the problem.
    • Separate test results from proven business impact, and explain remaining limits.

    What the interviewer checks: The interviewer checks whether you can explain your own work, justify choices, and assess results honestly.

    Other ways it was asked

    • “Can you describe a specific project where you used Python and machine learning frameworks to solve a real-world problem, including the problem, your approach, and the outcome?”
    • “Tell me about your experience working with machine learning or deep learning frameworks. What have you built? What challenges did you run into?”

Related interview questions

• Machine Learning interview questions
• Deep Learning interview questions
• Computer Vision interview questions
• MLOps interview questions
• Python interview questions
• Data Engineer interview questions
• Data Analyst interview questions
• Data Scientist interview questions
• AI Engineer interview questions
• AI Trainer interview questions

How this list is made

Every question here was asked by a real interviewer in a real job interview that a candidate ran MeetAssist in. After the call, the interviewer's questions are pulled from the transcript. Practice sessions, tests, exams and anything that was not a job interview are dropped.

Questions are rewritten so they make sense on their own: names of people, employers and clients are removed, and speech-to-text mistakes are fixed. Small talk, logistics and questions about one candidate's own résumé are left out. Rephrasings of the same question are merged, and the count is the number of different interviews it came up in.

The interviews come from candidates who use MeetAssist, so they lean towards software, data and AI roles: the shares describe these interviews, not every Machine Learning Engineer interview. Answers to the most asked questions are written and checked by an editor; a "common mistake" is shown only when reviews of real answers in these interviews show it, with the count.

69 interviews, 50 candidates, Dec 2025 to Oct 2026. Updated every month with new interviews.