AWS Certified Machine Learning Engineer – Associate — Practice Set 1

Question 1

A data engineer at a retail company needs to prepare a 500 GB dataset of customer purchase records stored in Amazon S3 for training a recommendation model. The dataset contains missing values in several numeric columns. The engineer wants to perform visual data profiling, apply transformations, and handle missing values with MINIMAL custom code. Which AWS service BEST meets these requirements?

Accepted Answer

AWS Glue DataBrew is a visual data preparation service that provides over 250 built-in transformations including missing value imputation (mean, median, mode, custom value), outlier handling, and data profiling — all through a no-code visual interface. It directly reads from and writes to Amazon S3. Option A requires writing and maintaining custom PySpark code. Option C requires writing scikit-learn scripts and managing SageMaker Processing job infrastructure. Option D requires two separate services and custom Lambda code to rewrite Parquet or CSV files.

Question 2

A machine learning engineer is preparing a dataset for a fraud detection model. The dataset has 1 million non-fraudulent transactions and only 5,000 fraudulent transactions, resulting in severe class imbalance. The engineer must address the imbalance before training. Which approach within Amazon SageMaker MOST effectively addresses class imbalance for a binary classification model?

Accepted Answer

The Amazon SageMaker built-in XGBoost algorithm's scale_pos_weight hyperparameter directly adjusts the weight given to the positive (minority) class during training, effectively compensating for class imbalance without requiring data resampling. With a 200:1 ratio, setting scale_pos_weight=200 makes the algorithm focus equally on both classes. This is the most efficient approach because it avoids generating synthetic data and requires only a hyperparameter change. Option A is valid but Amazon SageMaker Data Wrangler does not natively support SMOTE as a built-in transform (it requires a custom transform). Option B Autopilot handles some imbalance internally but gives the engineer less control. Option C is technically correct but adds the overhead of a Processing job when the algorithm itself can handle the imbalance.

Question 3

An organization needs to build a feature store that allows multiple ML teams to share and reuse features across different models. Features must be available both for low-latency online inference (sub-10ms) and for offline batch training. The solution must be fully managed. Which AWS service should a machine learning engineer use?

Accepted Answer

Amazon SageMaker Feature Store is a fully managed feature store that provides an online store for sub-millisecond retrieval during inference and an offline store (backed by Amazon S3) for batch training data. It includes a feature group catalogue, versioning, and point-in-time correct queries to prevent data leakage. Option A requires building and maintaining custom synchronisation pipelines between DynamoDB and S3. Option C ElastiCache lacks the feature catalogue, versioning, and point-in-time query capabilities needed for ML feature stores. Option D Amazon Redshift provides acceptable batch query performance but cannot deliver sub-10ms online feature retrieval for real-time inference.

Question 4

A machine learning engineer is building a training pipeline for a natural language processing (NLP) model. The raw text data is stored in Amazon S3 and must be tokenised, padded to a fixed sequence length, and converted to TFRecord format before training. The processing script uses TensorFlow and requires 32 GB of memory. The engineer wants to run this as a managed, scalable job. Which AWS service BEST meets this requirement?

Accepted Answer

Amazon SageMaker Processing jobs run managed, containerised scripts that can use pre-built or custom containers (including TensorFlow) on configurable instance types. The ml.m5.4xlarge provides 64 GB of RAM, meeting the 32 GB requirement with headroom. Processing jobs auto-terminate after completion and require no persistent infrastructure. Option A AWS Lambda has a maximum memory limit of 10 GB, which is insufficient for this 32 GB requirement. Option C AWS Glue ETL is optimised for Spark-based data transformation and does not easily support heavy TensorFlow workloads or custom ML library containers. Option D Amazon EMR Serverless can run Python UDFs but is not designed for containerised TensorFlow processing and adds Spark overhead unnecessarily.

Question 5

A machine learning engineer is preparing a dataset for a time-series forecasting model. The dataset contains daily sales records from 2015 to 2025 for 10,000 product SKUs stored in Amazon S3. The engineer must engineer lag features (7-day, 30-day, 90-day lags), rolling averages, and product-level encoding without leaking future data into past windows. The job must complete within 2 hours and cost under $50. Which approach BEST meets all constraints?

Accepted Answer

AWS Glue ETL with PySpark is well-suited for large-scale tabular feature engineering. Spark window functions (partitionBy('sku').orderBy('date')) correctly compute lag values and rolling averages per SKU without leaking future data, since the ordering ensures past-only window frames. Using 20 G.1X workers provides 80 vCPUs and 160 GB RAM distributed across the cluster, sufficient to process 10 years of daily data for 10,000 SKUs well within 2 hours. At roughly $0.44/DPU-hour, 20 workers for 2 hours costs approximately $35, under the $50 budget. Option A SageMaker Data Wrangler has limited support for window-based lag features across thousands of SKUs and exports can be slow at this scale. Option B uses a single ml.m5.4xlarge (16 vCPUs, 64 GB) which may be insufficient for 10,000 SKUs over 10 years within 2 hours. Option D Autopilot generates features automatically but gives the engineer no control over specific lag windows or guarantee of no data leakage.

Question 6

A company collects raw images from retail stores and stores them in Amazon S3. A machine learning engineer must build a labelling workflow so that human reviewers can annotate the images with bounding boxes around product categories. The workflow must support quality control through annotation consensus across multiple reviewers. Which AWS service BEST supports this use case?

Accepted Answer

Amazon SageMaker Ground Truth is purpose-built for creating labelled training datasets. It supports bounding box annotation tasks, multiple workforce options (Mechanical Turk, private, or vendor), and annotation consolidation algorithms that compute consensus across multiple annotators — directly providing quality control. Option A Amazon Rekognition Custom Labels requires pre-labelled training data; it cannot label unlabelled images without an existing labelled dataset. Option C AWS Glue DataBrew is a data preparation service for tabular data, not image labelling. Option D Amazon A2I is designed for human review of ML model predictions (active learning loop), not for initial annotation of raw unlabelled data.

Question 7

A machine learning engineer needs to split a large labelled dataset stored in Amazon S3 into training (70%), validation (15%), and test (15%) sets in a reproducible and stratified way, ensuring that the class distribution is preserved across all three splits. Which approach requires the LEAST operational overhead?

Accepted Answer

Amazon SageMaker Data Wrangler provides a built-in Split Data transform that supports stratified splitting, preserving class distribution across splits, with configurable ratios. The output can be exported directly to Amazon S3 without writing any code. Option B achieves the same result but requires provisioning an EC2 instance and writing, running, and maintaining a Python script. Option C Autopilot determines its own internal split and does not expose stratified dataset splits for the engineer to use externally. Option D AWS Glue ETL's random split does not support stratification by class label.

Question 8

A machine learning engineer needs to train an image classification model to identify defective products on a manufacturing line. The company has 2,000 labelled images and limited ML expertise. The engineer needs a solution that trains a high-accuracy model with MINIMAL code and automatically tunes hyperparameters. Which Amazon SageMaker capability BEST meets this requirement?

Accepted Answer

Amazon Rekognition Custom Labels allows teams with no ML expertise to upload labelled images, click Train, and get a custom image classification or object detection model — with no code required and automatic model training. It is ideal for 2,000 images and industrial quality inspection use cases. Option A SageMaker Autopilot supports tabular data (CSV/Parquet) for classification and regression, not image data. Option C SageMaker JumpStart fine-tuning requires writing Python code or using Studio notebooks and basic ML knowledge. Option D requires manual hyperparameter tuning expertise and infrastructure knowledge.

Question 9

A machine learning engineer is training a deep learning model on Amazon SageMaker using a single ml.p3.8xlarge instance with four V100 GPUs. The training job takes 48 hours. The engineer wants to reduce training time to under 12 hours by scaling to multiple instances while keeping costs reasonable. Which Amazon SageMaker training feature should the engineer use?

Accepted Answer

The SageMaker distributed data parallel (SMDDP) library implements AllReduce-based data parallelism optimised for AWS network infrastructure, scaling training across multiple instances with near-linear speedup. Using 4 x ml.p3.8xlarge instances (16 GPUs total, 4x the original) should reduce training from 48 hours to approximately 12 hours. Option A Spot Training reduces cost by using spare capacity but does not inherently reduce training time; the job may also be interrupted. Option C Automatic Model Tuning runs parallel trials to find the best hyperparameters, not to speed up a single training run. Option D SageMaker Debugger identifies performance bottlenecks but does not itself distribute training across multiple instances.

Question 10

A data scientist has trained a custom PyTorch sentiment analysis model in an Amazon SageMaker Studio notebook. The model achieves acceptable accuracy, but the engineer wants to experiment with different learning rates, batch sizes, and dropout rates to optimise performance. The engineer wants the experiments tracked automatically, including metrics, parameters, and artefacts, without changing the training script significantly. Which Amazon SageMaker feature BEST supports this requirement?

Accepted Answer

Amazon SageMaker Experiments provides a tracking system for ML experiments, automatically recording hyperparameters, metrics, and artefacts for each trial. The SageMaker Python SDK's Tracker API requires only a few lines of code added to the training script. Engineers can compare trials in SageMaker Studio's Experiments UI or programmatically. Option A SageMaker Automatic Model Tuning automates the search process but requires launching separate training jobs for each configuration and does not provide the rich experiment tracking UI that SageMaker Experiments does. Option C SageMaker Debugger captures tensor-level debugging information and training metrics but is not designed as an experiment comparison framework. Option D SageMaker Model Monitor tracks deployed model behaviour over time, not training experiment results.

Question 11

A financial services company is training a gradient boosting model to predict loan default risk. The model must be interpretable to satisfy regulatory requirements — specifically, the company must be able to explain the top contributing features for each individual prediction. The model will be trained on Amazon SageMaker and deployed for batch inference. Which combination of Amazon SageMaker features BEST satisfies the interpretability requirement?

Accepted Answer

Amazon SageMaker Clarify computes SHAP values for individual predictions, providing per-sample feature attributions that explain exactly how much each feature contributed to a specific loan default prediction. This satisfies regulatory requirements for individual prediction explainability. It integrates directly with SageMaker Batch Transform, adding SHAP values to the output payload. Option B SageMaker Debugger captures per-iteration feature importance during training (average feature importance), not per-prediction SHAP values for individual samples. Option C Autopilot's explainability reports provide aggregate feature importance at the model level, not per-prediction explanations required by regulators. Option D SageMaker Model Monitor detects distribution drift in inputs and outputs but does not compute feature attributions for individual predictions.

Question 12

A machine learning engineer is evaluating a binary classification model trained on Amazon SageMaker. The model will be used to identify high-risk patients for a preventive health programme. False negatives (missing high-risk patients) are much more costly than false positives. Which metric should the engineer PRIMARILY optimise when selecting the final model threshold?

Accepted Answer

Recall (Sensitivity) = True Positives / (True Positives + False Negatives). When false negatives are the most costly error — missing a high-risk patient who then does not receive preventive care — maximising recall directly minimises this cost by ensuring as many actual high-risk patients as possible are flagged. Option A Accuracy is misleading with imbalanced classes and does not differentiate between the costs of false positives and false negatives. Option B Precision focuses on minimising false positives (flagging low-risk patients as high-risk), which is the less costly error in this scenario. Option D Specificity measures true negative rate for the negative class, which is also less relevant when false negatives carry the highest cost.

Question 13

A company has trained a large language model (LLM) foundation model and wants to fine-tune it on proprietary customer service conversation data stored in Amazon S3. The fine-tuning dataset contains 50,000 conversation pairs. The company wants to use a parameter-efficient fine-tuning technique to reduce GPU memory usage and training time while maintaining model quality. Which approach on Amazon SageMaker BEST meets these requirements?

Accepted Answer

Amazon SageMaker JumpStart supports parameter-efficient fine-tuning (PEFT) techniques including LoRA, which adds small trainable rank decomposition matrices to the model's attention layers while freezing the original weights. This dramatically reduces GPU memory usage (training only 0.1–1% of parameters) and training time while achieving comparable performance to full fine-tuning. The ml.g5.12xlarge provides 4 A10G GPUs suitable for this workload. Option B Amazon Bedrock's fine-tuning API performs full fine-tuning (all weights) rather than parameter-efficient fine-tuning, which requires more GPU resources. Option C Automatic Model Tuning searches hyperparameter space and does not implement LoRA or other PEFT techniques. Option D standard transfer learning fine-tunes all model weights, not a parameter-efficient subset.

Question 14

A machine learning engineer has trained a scikit-learn classification model on Amazon SageMaker and needs to deploy it for real-time predictions. The model will receive up to 100 requests per second with a latency requirement of under 200ms. The engineer wants a managed endpoint that automatically scales with traffic. Which deployment option BEST meets these requirements?

Accepted Answer

Amazon SageMaker real-time inference endpoints provide persistent, low-latency serving infrastructure for models with consistent traffic. Configuring Application Auto Scaling based on the InvocationsPerInstance CloudWatch metric allows the endpoint to scale out (add instances) as traffic increases beyond 100 req/s. Real-time endpoints consistently achieve sub-200ms latency for scikit-learn models. Option B Batch Transform is for offline batch inference, not real-time online predictions with latency requirements. Option C AWS Lambda deployment is limited by the deployment package size (10 GB), cold start latency, and is less suitable for CPU-intensive model inference at 100 req/s. Option D SageMaker Serverless Inference is cost-effective for intermittent traffic but can have cold starts that exceed 200ms for infrequent patterns.

Question 15

A machine learning engineer needs to automate a monthly retraining pipeline for a customer churn prediction model. The pipeline must: (1) extract and preprocess new training data from Amazon S3, (2) train a new model version using Amazon SageMaker, (3) evaluate the model against a holdout set, and (4) only register the new model in the SageMaker Model Registry if its AUC exceeds the currently deployed model's AUC. Which AWS service BEST orchestrates this pipeline?

Accepted Answer

Amazon SageMaker Pipelines is purpose-built for ML workflow orchestration and natively integrates all required steps: SageMaker Processing for data preparation, SageMaker Training for model training, and a Condition step that compares AUC metrics and conditionally registers the model in the SageMaker Model Registry. Amazon EventBridge Scheduler can trigger the pipeline on a monthly cron schedule. This approach requires no external orchestration and maintains full ML lineage and artefact tracking. Option B Step Functions can orchestrate these steps but requires more custom integration code and does not natively track ML lineage and model registry integration. Option C MWAA provides excellent DAG orchestration but introduces a managed Airflow cluster and requires Airflow-specific provider knowledge. Option D AWS CodePipeline is designed for CI/CD of software, not ML pipeline orchestration with conditional model registration.

Question 16

A company wants to deploy two versions of a recommendation model — a new candidate model and the current production model — and gradually shift 10% of live traffic to the candidate model to validate its performance before full rollout. Which Amazon SageMaker deployment strategy enables this traffic splitting?

Accepted Answer

Amazon SageMaker supports multiple production variants on a single endpoint, each with configurable traffic weights. Setting InitialVariantWeight to 0.1 for the new model and 0.9 for the existing model directs exactly 10% of requests to the candidate and 90% to the production model. The weight can be adjusted incrementally without redeploying the endpoint. Option A using Route 53 weighted routing adds unnecessary network complexity and DNS-based routing has coarser granularity. Option B multi-model endpoints share a single container and routing is done internally, not based on configurable traffic weights between model versions. Option D blue/green deployment is a complete cutover strategy, not gradual traffic splitting.

Question 17

A company needs to run batch inference on 10 million product images stored in Amazon S3 to classify them into 500 categories using a fine-tuned ResNet model. The inference job runs once per week and must complete within 8 hours. The model artefact is 2 GB. Cost must be MINIMISED. Which Amazon SageMaker feature and instance strategy MOST cost-effectively meets these requirements?

Accepted Answer

SageMaker Batch Transform is designed for offline large-scale inference jobs and automatically distributes the input dataset across multiple instances without a persistent endpoint. Using Amazon EC2 Spot Instances through SageMaker Managed Spot Training reduces compute costs by up to 90% compared to On-Demand. Multiple ml.g4dn.xlarge instances (each with an NVIDIA T4 GPU) parallelize the 10 million image workload. Checkpointing ensures that if a Spot Instance is interrupted, the job resumes from the last checkpoint rather than starting over. Option A real-time endpoints require persistent running instances even between weekly jobs, wasting cost. Option C Asynchronous Inference is designed for variable-latency requests, not the most cost-effective for scheduled weekly batch processing of 10M items. Option D a single instance, even a large one, cannot process 10 million images in 8 hours and is On-Demand pricing (expensive).

Question 18

A machine learning engineer needs to deploy the same trained Amazon SageMaker model to edge devices in retail stores that have limited internet connectivity and 4 GB of RAM. The model must run locally on the edge device without calling back to AWS for inference. Which AWS service should the engineer use to compile and package the model for edge deployment?

Accepted Answer

Amazon SageMaker Neo compiles trained models (from TensorFlow, PyTorch, XGBoost, and others) for specific hardware targets (ARM, x86, NVIDIA Jetson, etc.), producing an optimised binary that runs efficiently within the device's RAM constraints. AWS IoT Greengrass provides the runtime and deployment mechanism to push the compiled model and inference component to edge devices, running inference locally without requiring internet connectivity. Option B real-time endpoints require constant internet connectivity to reach the cloud, which violates the limited-connectivity constraint. Option C Lambda@Edge runs at CloudFront PoPs and requires the edge device to call out over the internet for each inference. Option D Serverless Inference also requires internet connectivity to reach the SageMaker endpoint.

Question 19

A machine learning engineer has deployed a binary classification model on an Amazon SageMaker real-time endpoint. After two months in production, the model's accuracy has significantly degraded. The engineer suspects that the statistical distribution of the input features has changed since the model was trained. Which Amazon SageMaker feature should the engineer enable to detect this automatically?

Accepted Answer

Amazon SageMaker Model Monitor's data quality monitor captures a statistical baseline (feature distributions, data types, value ranges) from the training data and continuously compares incoming live inference data against this baseline using statistical tests. When drift is detected — meaning input distributions have shifted — it generates CloudWatch metrics and alerts. This directly identifies data drift, which is the suspected cause of accuracy degradation. Option A SageMaker Debugger captures internal model tensors during training, not during deployed inference. Option C SageMaker Clarify computes SHAP values for bias detection, not for data distribution drift monitoring. Option D CloudWatch endpoint metrics track infrastructure performance (latency, throughput) but do not analyse input feature distributions.

Question 20

A company deploys a credit scoring model on an Amazon SageMaker endpoint. The compliance team requires that the model's predictions be continuously monitored for bias against protected demographic groups (age, gender) in production. The team needs automated alerts when bias metrics exceed acceptable thresholds. Which combination of Amazon SageMaker features BEST meets this requirement?

Accepted Answer

Amazon SageMaker Clarify bias monitoring integrates with SageMaker Model Monitor to compute continuous bias metrics (such as Disparate Impact, Equal Opportunity Difference) against a baseline for specified protected attributes like age and gender. The bias metrics are emitted as CloudWatch metrics, and CloudWatch alarms can trigger Amazon SNS notifications when thresholds are breached. Option A Model Monitor's model quality monitor measures prediction accuracy metrics (F1, precision, recall) but does not compute demographic bias metrics. Option C Experiments is for tracking training experiments, not for continuous production monitoring. Option D Debugger's class imbalance rule detects imbalance in the training dataset, not demographic bias in live production predictions.

Question 21

A company's ML team trains models in Amazon SageMaker Studio. The security team requires that all training data in Amazon S3 and model artefacts be encrypted using a company-managed AWS KMS key. The security team also requires that training jobs run in a VPC with no internet access. Which combination of SageMaker training job configuration options enforces both requirements?

Accepted Answer

Setting VolumeKmsKeyId to the company's CMK ARN encrypts the EBS volumes attached to the training instances (where data is copied and model artefacts are stored) using the company-managed key. Setting NetworkIsolation=True prevents the training container from making any outbound internet calls. Specifying a VPC configuration with private subnets places the training instances inside the company VPC. Together these satisfy both the KMS encryption and no-internet-access requirements. Option B SSE-S3 uses AWS-managed keys, not the company's CMK, and an S3 VPC endpoint alone does not block all internet access from the training container. Option C AWS Glue encryption is for Glue jobs, not SageMaker training volumes; Studio VPC-only domain restricts data science access but not the training job's network access. Option D SageMaker built-in encryption refers to S3-side artefact storage; AWS PrivateLink secures API calls but does not prevent the training container from accessing the internet.

Question 22

A machine learning engineer observes that an Amazon SageMaker real-time endpoint's latency has increased from 50ms to 800ms over three weeks, even though traffic patterns have not changed significantly. The endpoint runs a TensorFlow model on ml.c5.2xlarge instances. Amazon CloudWatch metrics show CPU utilisation consistently at 95% across all instances. The auto scaling policy is configured to scale out when CPUUtilization exceeds 70%. The engineer notices that new instances are added but removed within minutes due to the scale-in policy. What is the MOST likely cause and the appropriate fix?

Accepted Answer

With CPU at 95% on all instances even when traffic has not changed, the current instance size cannot handle the steady-state load — the auto scaling threshold is set at 70% but the baseline load already drives CPU to 95%. New instances added during scale-out reduce the per-instance load briefly, but the scale-in policy removes them too quickly, and CPU immediately spikes back to 95%. The root cause is undersized instances for the baseline workload, not a cooldown issue. Upgrading to ml.c5.4xlarge (8 vCPUs vs 4) doubles per-instance CPU capacity, bringing baseline utilisation to approximately 47%, well below the 70% threshold. Increasing the minimum instance count ensures the improved fleet size persists without scale-in. Option B addresses the cooldown symptom but not the root cause of undersized instances. Option A a memory leak would manifest as OOM errors or gradual memory growth, not sustained 95% CPU. Option D load balancer routing is managed by SageMaker and uneven routing would show high variance in per-instance metrics, not consistently high CPU across all instances.

Question 23

A company uses Amazon SageMaker to train and deploy ML models. The MLOps team wants to automatically trigger a retraining pipeline when model quality degrades below an AUC threshold of 0.80 in production. The model quality is monitored using Amazon SageMaker Model Monitor. Which combination of AWS services should the team use to automate the retraining trigger?

Accepted Answer

Amazon SageMaker Model Monitor with a model quality monitor emits custom CloudWatch metrics including AUC. A CloudWatch Alarm set on this metric triggers when AUC falls below 0.80. Configuring the alarm action to send an event to Amazon EventBridge, which then starts an Amazon SageMaker Pipelines execution, creates a fully automated retraining loop with no custom polling logic. Option B SageMaker Debugger is for training-time monitoring, not production endpoint quality monitoring. Option C a daily Processing job introduces up to 24 hours of delay before retraining is triggered and requires custom code to evaluate and act on the AUC metric. Option D bias monitoring tracks demographic fairness metrics, not model accuracy metrics like AUC.

Question 24

A machine learning team stores trained model artefacts in Amazon S3 and registers them in the Amazon SageMaker Model Registry. The team wants to track which version of the training dataset, which training job, and which hyperparameters were used to produce each registered model version. Which feature of Amazon SageMaker provides this end-to-end lineage tracking automatically?

Accepted Answer

Amazon SageMaker ML Lineage Tracking automatically records and stores the complete lineage graph connecting training datasets, processing jobs, training jobs, model artefacts, and deployed endpoints. Engineers can query the lineage graph to determine which exact dataset version and hyperparameters produced a specific registered model version, satisfying reproducibility and audit requirements. Option A SageMaker Model Monitor tracks live inference data quality, not training lineage. Option C AWS CloudTrail records API calls as audit logs but does not build a queryable lineage graph connecting entities across the ML lifecycle. Option D SageMaker Experiments tracks hyperparameters and metrics per training run but does not link to downstream registered model versions and deployed endpoints in a lineage graph.

Question 25

A company's Amazon SageMaker Studio domain is configured without a VPC. The security team mandates that all Studio notebook traffic to Amazon S3, Amazon SageMaker APIs, and other AWS services must route through the company VPC and never traverse the public internet. The team also requires that the Studio EFS volume containing notebook files be encrypted with a company CMK. Which configuration changes are required? (Choose TWO.)

Accepted Answer

SageMaker Studio's VPC mode must be set at domain creation time and cannot be changed on an existing domain — the domain must be recreated with VPC-only mode specified, which routes all Studio traffic through private subnets. VPC endpoints (AWS PrivateLink) for S3 and SageMaker APIs ensure traffic never leaves the AWS network. Separately, the EFS volume encryption key (KmsKeyId) for the Studio domain must also be specified at creation time. Both settings together address the network isolation and encryption requirements. Option B IAM condition-based VPC restrictions apply to API calls from notebooks but do not prevent the Studio interface itself from routing non-API traffic over the internet. Option D VPC-only mode cannot be changed on an existing domain; also, an internet gateway does not block outbound traffic by default. Option E PrivateLink cannot be enabled on an existing non-VPC domain without recreating it, and S3 bucket policies alone do not enforce Studio's routing.

MLA-C01 — AWS Certified Machine Learning Engineer – Associate