AWS Certified Machine Learning Engineer Associate: Hands On!

Explore data ingestion and storage fundamentals, including data warehouses, data lakes, data lake houses, and ETL pipelines, plus AWS storage services like S3, EBS, EFS, FSx, Kinesis, and Kafka.

Identify the three types of data—structured, unstructured, and semi-structured—and reference examples like databases, CSV files, XML and JSON, and logs to understand queryability.

Explore the three data v's—volume, velocity, and variety—and learn how data size, speed, and formats shape storage, processing, and unified queries.

Compare data warehouses, data lakes, and data lake houses, and learn schema on write versus schema on read, ETL versus ELT, with AWS Redshift and S3.

Explore data mesh as a decentralized data governance model where teams own data and offer data products, with federated governance and central standards, often implemented with AWS lake formation.

Explore ETL and ELT pipelines, detailing extraction from diverse data sources, transforming data to cleanse and enrich, and loading into data warehouses or data lakes, with orchestration via AWS tools.

Explore data sources such as JDBC and ODBC, raw log files, APIs, and real-time streams. Learn common formats like CSV, JSON, Avro, and Parquet and how they power data pipelines.

Explore Amazon S3, the backbone of cloud storage, with backup, disaster recovery, and archival use cases. Learn about buckets, region, object keys and prefixes, with versioning and metadata.

Create a general-purpose S3 bucket in a chosen region with account regional namespace, configure security, upload objects, organize with folders, and understand pre-signed URLs and public access.

Implement Amazon S3 security using user-based and bucket policies, cross-account access, public reads, encryption, and JSON policies.

Enable public access for an S3 bucket by creating a public bucket policy with the policy generator, granting getObject to all objects via the bucket arn.

Enable bucket-level versioning in Amazon S3 to create file versions on upload, use delete markers to recover from deletions, and roll back to prior versions.

Enable S3 bucket versioning to track changes, create new versions on overwrite, and learn to roll back by deleting a specific version or delete marker.

Explore Amazon S3 replication with crr and srr, enabling asynchronous cross-region or same-region replication between buckets, requiring versioning and proper iam permissions, and covering practical use cases.

Enable Amazon S3 replication to automatically replicate new objects, use S3 batch replication for existing and failed items, and replicate delete markers from source to target without chaining.

Set up Amazon S3 cross-region replication by creating origin and replica buckets, enabling versioning, mirroring new objects and their versions, and enabling delete marker replication.

Explore the full range of Amazon S3 storage classes, from Standard and Infrequent Access to Glacier and Intelligent Tiering, and learn durability, availability, and lifecycle automatic tiering.

Demonstrates creating a S3 bucket, uploading an object, and exploring storage classes from Standard to Glacier tiers—including Standard-IA and One-Zone-IA—plus automating transitions with lifecycle rules.

Move objects between storage classes with lifecycle rules—transitions from standard to standard IA, intelligent tiering, one-zone IA, glacier, and deep archive—plus expiration, versioning, S3 analytics, and prefix or tag filters.

Create a lifecycle rule to move current and non-current versions across storage classes. Expire and permanently delete old data, and clean up delete markers or incomplete uploads.

Explore how Amazon S3 event notifications trigger on object events, filter by criteria like JPEG, and deliver to SNS, SQS, Lambda, or Event Bridge with versatile destinations.

This hands-on lecture demonstrates configuring S3 event notifications to publish ObjectCreated events to SQS, with optional EventBridge, Lambda, or SNS destinations, and tests using a sample coffee.jpg.

Explore S3 baseline performance, with 3,500 put/copy/post/delete per second per prefix and 5,500 get/head per second per prefix. Optimize transfers with multi-part uploads, S3 transfer acceleration, and byte range fetches.

Explore Amazon S3 object encryption using SSE-S3, SSE-KMS, SSE-C, and client-side encryption with headers like x-amz-server-side-encryption and AES-256, plus encryption in transit via https.

Learn to configure AWS S3 default encryption with SSE-S3, SSE-KMS, and DSSE-KMS, enable bucket versioning, and manage KMS keys and object versions for secure storage.

Default encryption applies to new S3 objects with SSE-S3, and bucket policies evaluated before default encryption can require SSE-KMS or SSE-C by denying unencrypted puts.

Use s3 access points to simplify security management with per-point policies granting read/write access for finance and sales, and read-only access for analytics, each with dns name and vpc origin.

Discover how S3 object lambda uses access points and lambda to redact or enrich objects on retrieval, enabling pii redaction for analytics and enrichment for marketing from a single bucket.

Explore Amazon Elastic Block Store (EBS) volumes as network storage for EC2 instances, enabling data persistence across terminations, availability zones, snapshots, and configurable capacity and IOPS.

Learn to manage Amazon EBS volumes for an EC2 instance by creating, attaching, and deleting GP2 volumes across availability zones, and review delete on termination for root volumes.

Discover EBS elastic volumes that let you modify volumes on the fly via the console, increasing size and changing type (GP2 to GP3) without downtime, and set IOPS or throughput.

Explore Amazon EFS, a scalable, pay-as-you-go NFS that spans multiple EC2 instances and zones, with encryption at rest, security groups, and flexible throughput and storage classes.

Set up a regional amazon elastic file system (efs) and mount it to two ec2 instances across availability zones, demonstrating backups, lifecycle, and elastic throughput for shared storage.

EBS volumes attach to a single instance and stay AZ-bound; use snapshots for AZ migration and IO control, while EFS remains a shared cross-AZ file system with storage tiers.

Explore amazon fsx, a fully managed service that launches four file systems on AWS—Windows File Server, Lustre, NetApp ONTAP, and OpenZFS—for machine learning and high-performance computing workloads with S3 integration.

Explore Amazon FSx file systems: Lustre, Windows File Server, NetApp ONTAP, and OpenZFS, with high-level choices on throughput, storage type, encryption, and AZ deployment for exam readiness.

Explore Amazon kinesis data streams for real-time data ingestion with producers and consumers, shard-based capacity, retention for replay, security, and on-demand or provisioned modes, using KPL, KCL, and Firehose.

Conduct a hands-on walkthrough of creating and using a Kinesis Data Stream, sending and reading records via the AWS CLI, and exploring on-demand versus provisioned capacity, shards, and producer/consumer options.

Learn how Amazon Data Firehose streams data from producers to destinations with optional Lambda transformations and buffering for near real-time delivery to S3, Redshift, OpenSearch, or third-party endpoints.

Troubleshoot Kinesis data streams by diagnosing producer bottlenecks, throughput limits, and hot shards. Apply batching, partition key tuning, and exponential backoff, then monitor consumer latency and shard scaling.

Explore the shift from Kinesis Data Analytics to the managed service for Apache Flink, enabling SQL and Table API processing of streaming data with S3 reference tables.

Discover how to deploy real-time data processing with the managed service for Apache Flink, using Kinesis streams to an S3 destination, via blueprint or streaming app workflows.

Explore the cost model of kinesis analytics, a serverless service that charges for resources used and scales automatically, with IAM security, schema discovery, and random cut forest for anomaly detection.

Learn how AWS MSK, a fully managed Apache Kafka service on AWS, deploys multi-AZ Kafka clusters in a VPC and compares to Kinesis for data streams.

Deploy MSK Connect to run managed Kafka Connect workers with auto-scaling on AWS, and attach connectors like Amazon S3 or OpenSearch to move data from your MSK cluster to destinations.

Explore MSK serverless: Apache Kafka runs without capacity management, automatically scales compute and storage as you define topics and partitions; includes IAM access control and per-cluster and per-partition pricing.

Compare Kinesis data streams and Amazon MSK, focusing on 10 megabytes per message versus 1 megabyte by default, and shard versus partition scaling.

Explore data transformation, integrity, and feature engineering, including handling missing data, outliers, and imbalanced data, with hands-on labs on AWS EMR, SageMaker Data Wrangler, SageMaker Clarify, and Glue.

Explore elastic MapReduce (EMR) as a managed Hadoop framework on EC2, delivering distributed data processing with Spark, HBase, Presto, Flink, and Hive, plus the EMR notebook and S3-backed storage.

Explore EMR serverless, which automatically provisions capacity for Spark, Hive, or Presto jobs via S3-based scripts, while you can pre-initialize and adjust resources as needed.

Explore the Hadoop ecosystem with HDFS, Yarn, MapReduce, and Hadoop core running on EMR. See how Spark on EMR delivers in-memory processing, Spark SQL, MLlib, streaming, and graph processing.

Master feature engineering for machine learning by selecting and transforming features to mitigate the curse of dimensionality, using normalization, missing data handling, and dimensionality reduction techniques like PCA and k-means.

Prepare data at scale with Spark and EMR Studio to compute tf-idf scores, using term frequency and inverse document frequency across unigrams and bigrams for search relevance.

Recreate a wikipedia-like dataset using tf-idf with spark on emr studio's serverless notebook, performing tokenization, hashing tf, and idf in PySpark to rank documents.

Learn how to impute missing data in feature engineering, including mean and median imputation, with cautions about outliers and bias, and explore advanced methods such as KNN, MICE, and regression.

Learn to handle unbalanced data in fraud detection, using oversampling, undersampling, and SMOTE, and adjust decision thresholds to balance false positives and negatives.

Learn how to identify and handle outliers using variance, standard deviation, and box-and-whisker concepts, and explore when to remove them, with AWS Random Cut Forest as a detection tool.

Explore feature engineering techniques such as binning, quantile binning, transforming features, one-hot encoding, scaling and normalization, and data shuffling to improve model input quality and capture non-linear relationships.

Explore SageMaker as a comprehensive platform for the end-to-end machine learning lifecycle, from data preparation and training to deployment via endpoints, using notebooks, S3, and ECR.

Discover how a SageMaker AI domain unifies users, applications, and shared resources around a central efs volume and vpc settings, with private user workspaces and SageMaker studio access.

Prepare data in SageMaker from S3 or FSx for Lustre, train with built-in or custom algorithms, and deploy via runtime endpoints or batch transform with scalable inference.

SageMaker ground truth uses a human labeling workforce to generate labels and features, building a model over time to label only uncertain cases, and reduces labeling costs up to 70%.

Learn how Amazon Mechanical Turk serves as a crowdsourcing marketplace, assigning simple tasks to a distributed human workforce for image labeling and data collection, with fast rewards.

Explore SageMaker Data Wrangler, a code-generating ETL tool in SageMaker Studio that imports, visualizes, and transforms data with 300+ operations, runs quick models, and exports code for notebooks.

Explore SageMaker Studio and Canvas for end-to-end machine learning, using Data Wrangler for data wrangling, data quality insights, and quick modeling to predict absenteeism from a UCI repository data set.

Leverage SageMaker model monitor to detect data drift, quality issues, anomalies, and new features with CloudWatch alerts, visuals, and no-code setup, plus SageMaker Clarify for bias detection and feature attribution.

Explore how SageMaker Clarify uses partial dependence plots to show feature influence on model predictions and how Shapley additive explanations, including asymmetric Shapley values for time series, explain forecast behavior.

SageMaker feature store organizes features into feature groups and provides fast, secure online and offline stores for streaming and batch access.

Explore SageMaker canvas, a no-code machine learning environment for business analysts that uses AutoML on CSV data, auto cleans data, supports classification and regression, and enables export to SageMaker Studio.

Discover how AWS Glue is a serverless metadata hub for your data lake, extracting structure from S3, cataloging schemas, and enabling ETL and SQL queries with Athena, Redshift, and EMR.

Discover AWS Glue Studio, an ETL tool for building DAG workflows with no code, connecting sources like S3, Kinesis, Kafka, or JDBC, and exporting to S3 or Glue Data Catalog.

Learn how AWS Glue Data Quality automatically evaluates data with quality rules, logs results to CloudWatch, and integrates as a Glue transformation using either auto-generated or manually defined DCDL rules.

Explore AWS Glue DataBrew, a visual data preparation tool for pre-processing large datasets with drag-and-drop transformations (over 250 options) and output to S3.

Explore Glue DataBrew to preprocess data with 250 plus transformations, build recipes, filter and clean data, and run interactive or scheduled jobs without writing any code.

Explore how to handle personally identifiable information in DataBrew transformations using substitution, shuffling, deterministic and probabilistic encryption, masking, deletion, and hashing.

Explore Amazon Athena, a serverless sql interface for s3 data that queries csv, json, avro, parquet, and orc without loading data into a database.

Query S3 data via a Glue data catalog with Athena, a serverless SQL engine. Use work groups to manage access and costs, and optimize with columnar formats ORC or Parquet.

Create a table from a query using CTAS in Athena, converting data formats like parquet or ORC and storing the results in S3 with optional compression.

Improve Athena performance by using columnar formats such as Orc or parquet and preprocessing with Glue or ETL; prefer a few large files and MSK repair table to add partitions.

Athena's acid transactions guarantee safe concurrent row-level modifications and time travel using Apache Iceberg; enable this by creating an iceberg table in Athena or via lake formation governed tables.

Explore fine-grained access in Athena for the AWS Glue Data Catalog by mapping database and table operations to IAM actions, including drop, show, and create, across regions.

Explore AWS managed AI services that handle training and deployment for you, enabling image analysis, translation, text-to-speech, speech-to-text, and chatbots. Learn what each service offers for the exam.

Discover AWS managed services for pre-trained machine learning, covering GenAI with Bedrock and SageMaker, text and vision tools, chatbots, and scalable, cost-efficient, regionally available solutions.

Explore Amazon Comprehend, a fully managed natural language processing service that uses machine learning to extract key phrases, people, places, brands, topics, and sentiment, with custom classification and entity recognition.

Learn to build and manage Amazon Comprehend custom models for entity recognition and document classification, including versioning, cross-account sharing, and importing models via the Comprehend console.

Explore Amazon Comprehend to extract entities, PII, sentiment, and syntax from unstructured text, and train a custom classifier with S3 data for real-time categorization.

Explore Amazon Translate, a natural and accurate language translation service that localizes content for international users, including websites and apps, and translates large volumes of text efficiently.

Explore Amazon Translate, a neural network-based translation service, with document translation, batch translation, and source-to-target language options, plus custom terminology and parallel data for style control.

Amazon Transcribe converts speech to text using speech recognition, enabling PII redaction, language identification, and multilingual transcription; it supports custom vocabularies and language models to boost accuracy, and toxicity detection.

Explore Amazon Transcribe by streaming audio to text, redact PII, and enable multilingual transcription with automatic language identification for English and French.

Amazon Polly turns text into lifelike speech with deep learning, enabling apps to talk. Use lexicons, SSML, and voice engines (neural, standard, long-form, generative), speech marks for timing and lip-syncing.

Explore amazon polly's text-to-speech capabilities, compare generative, neural, and standard engines, and learn to use ssml markup, pauses, and pronunciation customization for natural speech outputs.

Explore Amazon Rekognition to analyze images and videos for objects, people, text, and scenes, including face detection and labeling; use custom labels to train logos and automate moderation with A2I.

Explore Amazon Rekognition capabilities such as label detection with object detections, custom labels, and confidence scores, plus image properties, moderation, and facial analysis.

Amazon Lex enables chatbots with text or voice interfaces for hotel bookings. It integrates with AWS Lambda, Amazon Connect, Comprehend, and Kendra to fulfill intents in multiple languages using slots.

Learn how Amazon Lex enables building chatbots and conversational AI, using traditional bots with intents, utterances, slots, and a visual builder to configure bookings like hotel reservations.

Explore Amazon Personalize, a fully managed ml service with recipes for real-time personalized recommendations. Read input data from s3 and use the api to deliver user-specific recommendations, not forecasting.

Explore Amazon Textract, which uses AI to extract text, handwriting, and data from scanned documents, including forms, tables, PDFs, and images. Use cases span invoices, medical records, and IDs.

Explore Amazon Textract by analyzing scanned documents to extract raw text, identify layout elements like titles and forms, run queries, and retrieve key-value pairs, tables, and ID data.

Explore Amazon Kendra, a fully managed machine-learning document search service that builds a knowledge index, extracts answers from text, PDF, HTML, and more, with natural language search and incremental learning.

Amazon augmented AI enables human review for low-confidence predictions. High-confidence results return instantly, and reviewed outputs feed back to improve models with risk-weighted scores stored in history.

Explore Amazon Augmented AI in SageMaker, review Rekognition and Textract predictions, and set up low-confidence-based human review workflows for content moderation.

Explore Amazon EC2 basics for AI, including GPU and Trainium Inferentia instances, Elastic Compute Cloud concepts, and cost-efficient training and inference with scalable virtual servers.

Explore choosing Amazon EC2 instance types for machine learning, using get advice to match deep learning inference with G5g and C7gn families, and review training vs inference options.

Discover Amazon Q Business, a fully managed gen-ai assistant trained on your internal data to answer questions, generate content, and automate tasks using data connectors, plugins, and guardrails.

Learn to set up Amazon Q Business, create a demo app, enable anonymous access, add an S3 knowledge base, index documents, and sync data for an internal knowledge chat.

Explore Amazon Q Apps, a no-code platform that creates Gen AI-powered apps from natural language using your company data via the Amazon Q Apps creator web UI.

Create an Amazon Q Apps app that generates a summary of customer feedback from a file and publish it to your library to analyze data with gen ai based apps.

Unsubscribe and remove the user, delete the application and Amazon Q business to avoid any future costs.

Explore Amazon Q Developer, a dual tool that answers AWS docs questions, lists account resources, and offers AI code companion for multi-language code with real-time suggestions and security scans.

Explore Amazon Q and Amazon Q Developer in aws, including the free and pro tiers, and use ai coding assistant, cli, and CloudShell to manage s3 buckets.

Explore SageMaker's built-in algorithms to quickly deploy common machine learning workloads at scale, using existing models for tasks like XGBoost, image recognition, NLP, and clustering.

Understand how SageMaker orchestrates the full machine learning workflow—from data preparation in S3 with notebooks and built-in models to training, deploying endpoints, and continuously refining models with new data.

Explore SageMaker input modes to optimize data flow into training instances, comparing S3 file mode, S3 fast file mode, and pipe mode, including FSx for Lustre and EFS considerations.

Explore linear learner in sageMaker for regression and classification. Use protobuf float32 or CSV input, file or pipe mode, and SGD with Adam or Adagrad, plus L1/L2 regularization.

Explore XGBoost in SageMaker, a fast gradient boosting decision-tree algorithm for classification and regression. Learn about open-source roots, input formats, key hyperparameters, and CPU versus GPU training options.

Explore LightGBM, a gradient boosting decision tree algorithm that forms ensembles for classification, regression, and ranking. Learn CPU-only training, memory considerations, and key hyperparameters like learning rate and leaves.

Explore SageMaker's seq2seq that maps token sequences to output tokens for translation, summarization, and speech-to-text, using record io protobuf inputs and a vocabulary file, and train on GPU instances.

Learn how DeepAR in SageMaker forecasts multiple interrelated time series and captures seasonality. Use JSON lines or parquet inputs with dynamic and categorical features, training on the full data.

Explore blazing text for supervised text classification and word two vec embeddings, training with labeled sentences and leveraging cbow, skip gram, and batch skip gram modes on CPU or GPU.

Object2Vec in SageMaker creates low-dimensional embeddings for arbitrary objects, enabling nearest-neighbor search, clustering, and recommendations through two input channels with separate encoders and a comparator.

Explore object detection in sageMaker, identifying image objects with bounding boxes and confidence scores using MXNet SSD or TensorFlow models, with pre-trained options, training data formats, and GPU-based training.

Compare image classification with object detection, learning to label images without locating objects. Explore MXNet and TensorFlow implementations, transfer learning, fine-tuning top layers, and GPU-accelerated training and inference.

Learn how semantic segmentation produces pixel-level masks that label every pixel, enabling precise object mapping for applications like self-driving cars and medical imaging.

Explore Amazon's random cut forest, an unsupervised anomaly detection algorithm that scores anomalies in data series using a forest of trees; supports batch or streaming data with no training.

Explore unsupervised neural topic modeling in SageMaker to group documents into latent topics using neural variational inference, with tokenization, a vocabulary, and GPU or CPU training.

Explore SageMaker's latent dirichlet allocation for unsupervised topic modeling, producing unlabeled document groupings from word counts, with cpu based training and key hyperparameters like num_topics and AlphaZero.

Explore k-nearest neighbors, a simple method for classification and regression, enhanced by SageMaker with sampling, dimensionality reduction, and fast neighbor lookup; tune k and sample size on CPU or GPU.

Explore how k-means clustering in SageMaker performs unsupervised data partitioning into k clusters, with k-means++ initialization, elbow method, and scalable training on S3.

Principal component analysis is a dimensionality reduction technique that reduces high dimensional data to a lower space with covariance matrices and SVD, using SageMaker in regular or randomized unsupervised modes.

Explore factorization machines for classification or regression on sparse data, especially in recommender systems, using pairwise user-item interactions and matrix factorization to predict unseen ratings.

Identify anomalous activity with IP insights in SageMaker, an unsupervised neural network that learns entity and IP patterns to flag suspicious logins and resource creation.

Master deep learning foundations, neural networks, and model training, tuning, and evaluation for image recognition and time series data, with SageMaker tuning and monitoring tools.

Explore deep learning with convolutional and recurrent networks, tune topology and hyperparameters to reduce overfitting, and evaluate results using accuracy, precision, recall, f1, and ROC curves in SageMaker.

Explore activation functions inside neural networks, from linear and binary step to nonlinear options like sigmoid and tanh, and learn when to use softmax, ReLU family, and swish.

Explore convolutional neural networks and their feature location invariance, from 2D conv layers to maxpooling, dropout, flatten, and softmax in practical Keras and TensorFlow workflows.

Explore recurrent neural networks for sequences in time and language, including series and music. Learn memory cells, LSTM and GRU variants, and backpropagation through time for sequence to sequence tasks.

Explore tuning neural networks with gradient descent, focusing on learning rate and batch size across epochs to minimize a cost function and avoid local minima.

Explore regularization techniques to prevent overfitting in neural networks, including dropout and early stopping. Use training, evaluation, and testing data to gauge model performance and guide simplification, with CNN examples.

Compare L1 and L2 regularization, where L1 promotes feature selection by zeroing weights and producing sparse outputs, while L2 keeps all features with smaller or larger weights, enabling dense models.

Identify and address the vanishing and exploding gradient problems during neural network training. Explore remedies such as multi-level hierarchy, LSTM, ResNet, ReLU activations, and gradient checking.

Explore confusion matrices to understand true positives, true negatives, false positives, and false negatives in binary and multi-class predictions. Understand diagonal accuracy and how heat maps visualize per-class frequencies.

Explore how to read a confusion matrix and compute metrics like recall, precision, F1 score, specificity, RMSE, ROC and AUC, and precision-recall curves.

Move from classification metrics to regression metrics by using r squared, RMSE, and MAE to measure the accuracy of numerical predictions.

Explore ensemble learning through bagging and boosting, including random forests, parallelized training, and sequential reweighting. Boosting emphasizes accuracy, while bagging reduces overfitting and enables parallelization.

Learn how automatic model tuning in SageMaker optimizes hyperparameters efficiently by learning as it goes, exploring smart parameter ranges, parallel training, and best practices to control cost and improve accuracy.

Explore hyperparameter tuning in AMT with practical tips like early stopping and warm start to reduce compute time and avoid overfitting, using objective metrics after each epoch.

Explore SageMaker autopilot, an AutoML wrapper that automatically selects models, tunes hyperparameters, preprocesses data, and delivers optimized predictions with a notebook, leaderboard, deployment and monitoring.

Explore SageMaker Studio as a visual IDE for machine learning, offering collaborative notebook sharing and managed hardware, and use SageMaker Experiments to organize and compare historical ML jobs.

Master SageMaker debugger to capture training state and metrics, set automated rules and alerts, and use the SageMaker Studio Debugger dashboard with SME debug to profile TensorFlow, PyTorch, and MXNet.

Explore how SageMaker model registry serves as a centralized catalog for model metadata, versions, and approval status, enabling CI/CD, model sharing across applications, and production deployment with model cards.

Integrate TensorBoard with SageMaker to visualize training progress, including loss, accuracy, and model graphs. Explore histograms, embeddings projections, and profiling tools to optimize training and inspect changes over time.

Explore large-scale SageMaker training techniques, including the phased-out training compiler in deep learning containers, PyTorch optimization on GPU instances, and warm pools with keepalive periods and persistent cache.

Enable SageMaker checkpointing to create training snapshots for resume and debugging, synced to S3. Automatic cluster health checks replace faulty instances, restart jobs, and verify Nccl and GPU health.

SageMaker's distributed training libraries enable data parallelism and model parallelism with allreduce and allgather, using PyTorch DDP or other tools, after maxing out large single instances.

Explore SageMaker model parallelism library for PyTorch to train large models beyond single-GPU memory, using optimization state sharding, activation checkpointing, and activation offloading. Leverage v2 for easy setup.

Learn how the elastic fabric adapter (efa) and nccl accelerate distributed training on SageMaker, enabling trillion-parameter models through mix, data and model parallelism, and high-bandwidth p4d gpu instances.

Discover how transformer-based architectures drive large language models like GPT and Claude, and gain hands-on experience with SageMaker to see generative AI under the hood.

Explore the transformer architecture, self-attention, and positional encoding, tracing from RNN-based encoder-decoder models to parallelizable large language models and scalable training.

Explore how self-attention in transformers uses token embeddings, query, key, and value matrices, scaled dot-product attention with softmax, to produce context-aware token representations, with masking and multi-head setups.

Leverage transformers, foundation models with self-attention, to power chat, translation (code translation), sentiment analysis, named entity recognition, and summarization; they require fine-tuning, moderation, and careful testing to avoid incorrect results.

Discover how generative pre-trained transformers use decoder-only architectures, self-attention, token embeddings with tokenization, and sinusoidal positional encoding to generate text in parallel, trained by next-token prediction.

Explore how generative transformers produce the next token: from the final decoder output and embedding multiplication to probability distributions and temperature-driven randomness for diverse ideas.

Learn key LLM concepts like tokens, embeddings, and how top P, top K, and temperature control output randomness, plus context window and token limits.

Explore fine-tuning with transformers by starting from a pre-trained model, then train on task-specific data, freeze layers, or add top layers to tailor tone, classification, and responses.

Spin up a SageMaker notebook, install the Transformers package from Hugging Face, and explore tokenization, attention, and positional encoding with sine and cosine representations.

Visualize self-attention in Bert transformers with Bert vis and Hugging Face, exploring query, key, and value matrices and how context shapes attention heads.

Import a pre-trained hugging face model with the Transformers pipeline to generate text using GPT-2 in a SageMaker notebook.

Explore AWS foundation models, including Jurassic-2, Claude, and stable diffusion, and use SageMaker Jumpstart notebooks to load, feed, query, and fine-tune pretrained models for generation, embeddings, and search.

Explore SageMaker JumpStart foundation models with Hugging Face and deploy Falcon 40B Instruct BF16 in a SageMaker Studio. Learn to manage domains, notebooks, and resources while avoiding unexpected costs.

Build generative ai applications with Amazon Bedrock, exploring foundation models for text and image generation, knowledge bases, vector stores, and retrieval augmented generation, plus an llm agent with custom tools.

Amazon Bedrock, a serverless application programming interface unifying foundation models like Titan and Anthropic, with training, fine-tuning, retrieval augmented generation, and agent support.

Explore bedrock as a wrapper for diverse foundation models and experiment with chat, text, and image playgrounds while learning about prompts, model reasoning, tokens, and guardrails.

Fine tune bedrock models to create custom models trained on your data. Continue with continued pre-training using unlabeled data to reduce reliance on prompt engineering.

Explore retrieval-augmented generation with Bedrock by using a knowledge base and vector databases to augment prompts, improve relevance, and discuss pros and cons.

Explore how vector stores and embeddings power semantic search in Amazon Bedrock knowledge bases, using embedding vectors, vector databases, and retrieval augmented generation to retrieve top results.

Explore how Bedrock creates RAG knowledge bases by ingesting documents from S3, web crawlers, or third-party connectors, embedding them into vector stores for semantic retrieval.

Create a Bedrock knowledge base with a custom vector store, ingest data from S3, chunk text, embed vectors, and test retrieval with citations from queried chunks.

Learn how Amazon bedrock guardrails provide content filtering for prompts and responses in text foundation models, with word and topic filters, profanity and pii masking, and contextual grounding checks.

Create and test guardrails in Amazon Bedrock to block harmful content, enforce policies, and ensure responses are grounded across regions.

Build LLM agents on amazon bedrock by giving foundation models access to tools and data. Learn how planning, memory, rag, and knowledge bases enable agentic AI to act on requests.

Build a bedrock agent with a self-employment knowledge base and a weather action group guarded by guardrails. Configure memory, test traces, and deploy via alias while cleaning up resources.

Explore Bedrock features like importing your own models from SageMaker or S3, evaluating them with automatic metrics and human feedback, and using watermark detection and Bedrock Studio for collaboration.

Explore the inner workings of SageMaker, integrate your own containers, and optimize training and inference resources while covering MLOps essentials like ECR, CloudFormation, CDK, EventBridge, and Step Functions.

SageMaker deployment guardrails enable controlled model rollouts to asynchronous or real-time endpoints, using blue-green deployments with all-at-once, canary, and linear modes, shadow tests, monitoring, promotion to production, and automatic rollback.

Explore how SageMaker uses Docker containers in ECR to host training and inference across frameworks like TensorFlow and PyTorch, with production variants and variant weights for live A/B testing.

Deploy trained SageMaker models to edge devices with Neo, compiling TensorFlow, MXNet, PyTorch, Onnx, or XGBoost code for edge architectures. Pair Neo with IoT Greengrass to run inference locally.

Match SageMaker instance types to your algorithm, using gpu for deep learning and cpu for inference, and consider spot training with s3 checkpoints to save costs.

Learn how SageMaker automatic scaling uses scaling policies with target metrics and CloudWatch to adjust production inference variants, optimizing min/max capacity and cooldowns as traffic changes.

Deploy trained models to production with SageMaker by Jumpstart, the Python SDK, or CloudFormation, hosting real-time, asynchronous, or serverless inference endpoints with auto scaling and Neo optimization.

Explore SageMaker serverless inference and the inference recommender to automatically scale, monitor with CloudWatch, run load testing configurations, and optimize costs based on traffic, latency, and memory.

Learn to build inference pipelines by chaining 2–15 containers for real-time inference and batch transforms, combining pre-processing, predictions, and post-processing with Docker, Spark ML, and scikit-learn.

Explore how SageMaker model monitor (no code) automatically detects data drift, anomalies, and bias, visualizes baselines, and alerts via CloudWatch to safeguard model quality and transparency.

Capture inputs and inference outputs from SageMaker endpoints as json in S3 with model monitor data capture. Use the data for continuous training, debugging, and monitoring in real time.

Explore how to integrate SageMaker with Kubernetes and Kubeflow to build hybrid MLOps pipelines, leveraging SageMaker components, SageMaker projects, and SageMaker pipelines for processing, training, evaluation, and deployment.

Explore how Docker containers standardize app deployment across any OS, using Docker images stored in Docker Hub or Amazon ECR, with AWS ECS, EKS, and Fargate.

Learn how to deploy containers with Amazon ECS using EC2 launch type or Fargate, assign ECS task roles, integrate load balancers, and enable persistent shared storage with EFS.

Create your first Amazon ECS cluster and explore capacity options with Fargate, Fargate Spot, and an ASG provider; configure ECS instance roles, auto-scaling, and launch your first service.

Create an ECS service with a new task definition using Fargate, deploying nginxdemos-hello from Docker Hub. Configure an application load balancer and target group, then scale tasks.

Store and manage docker images on AWS with Elastic Container Registry, enabling private or public repositories, ECS integration, and IAM-protected access, plus vulnerability scanning, versioning, image tags, and lifecycle.

Explore Amazon EKS, AWS’s managed Kubernetes service, to run Kubernetes on AWS using EC2 or Fargate, with node options and storage class via CSI drivers.

Learn to create an Amazon EKS cluster, configure IAM roles and networking, and provision compute with managed node groups or Fargate profiles, plus add-ons like EBS CSI driver and EFS.

AWS Batch runs batch jobs using Docker images with dynamic provisioning of EC2 or spot instances, a serverless service that pays for underlying resources and can handle S3 cleanup tasks.

CloudFormation provides a declarative, infrastructure-as-code approach to deploying and managing AWS resources via templates, enabling automated resource creation, tagging, cost estimation, and cross-environment reuse.

Learn how CloudFormation uses templates to create and update EC2 instances, infrastructure as code. Explore change sets, elastic IPs, and security groups in us-east-1, with visualization in Application Composer.

Discover the AWS Cloud Development Kit (CDK) to define infrastructure with programming languages, synthesize CloudFormation templates, and deploy VPC, ECS Fargate services, ALB, Lambda, S3, Rekognition, and DynamoDB.

Deploy a full stack with the AWS CDK by creating an S3 bucket, a Lambda function that invokes Amazon Rekognition, and a DynamoDB table to store results.

Discover AWS CodeDeploy, a hybrid deployment service that upgrades apps from v1 to v2 on EC2 instances and on-premises servers via a single interface, with provisioning and CodeDeploy agent installation.

Discover how AWS CodeBuild builds code in the cloud by pulling from CodeCommit, executing your build script, and producing deployable artifacts for CodeDeploy.

Explore AWS CodePipeline, the managed ci/cd orchestration tool that connects CodeCommit and CodeBuild to build, test, and deploy code to servers such as Elastic Beanstalk, delivering fast updates.

Master git architecture and essential commands for version control, including clone, pull, push, branch and merge, plus stash, rebase, blame, and diff for exam readiness.

Compare git flow and GitHub flow to manage source code, detailing master, develop, feature, release, and hotfix branches, and explain when rapid deployments suit GitHub flow.

Leverage Amazon EventBridge to schedule cron jobs and react to events, routing to Lambda, SNS, SQS, or other services, with schema registry and cross-account policy support.

Explore Amazon EventBridge rules with event patterns and schedules to trigger on EC2 instance state changes and route events to targets like SNS, Lambda, or API destinations.

Explore how AWS Step Functions design and visualize workflows, manage errors and retries, audit histories, and time between steps for long-running processes, with examples like training models and batch jobs.

Explore AWS Step Functions, state machines, and state types (task, choice, wait, parallel, map, pass, succeed, fail) and how they orchestrate data pipelines on JSON, S3 objects, and CSV files.

Explore Amazon managed workflows for Apache Airflow, a hosted environment to run, visualize, and schedule Python-defined DAG workflows via S3, with VPC security and scalable orchestration.

Discover AWS Lake Formation, built on AWS Glue, to create a secure data lake on S3 with ETL, data catalogs, and access control, enabling queries via Athena, Redshift, and EMR.

Apply Lake Formation data filters to enforce column, row, or cell level security. Use the console or API to grant select permissions on a table via these filters.

apply the principle of least privilege by granting only the permissions needed to perform tasks. use an iam policy to restrict s3 access to a specific bucket and csv files.

Explore data masking and anonymization to protect sensitive information, including masking credit card numbers and passwords, using Redshift and Glue DataBrew policies, plus encryption, hashing, or safe deletion in etl.

Master SageMaker security with IAM and MFA, TLS, and CloudTrail, encrypt data at rest and in transit with KMS and S3, and enable inter-container traffic encryption under least-privilege access.

Secure SageMaker by configuring VPC endpoints for S3 access, using private VPC or Privatelink and NAT gateways. Control access with IAM policies and monitor via CloudWatch and CloudTrail.

Master identity and access management in AWS by creating users and groups, and applying IAM policies in JSON to enforce least privilege principle across EC2, Elastic Load Balancing, and CloudWatch.

Create IAM users and groups, and attach administrator access via a group policy. Sign in as an IAM user or via an account alias, avoiding root access.

Explore AWS console multi-session support to sign into multiple accounts from the same browser by adding sessions. See cross-account work in EC2 and EBS across separate windows for AWS usage.

Demonstrate how IAM policies attach to groups or users, enabling inherited or separate permissions, and explain the policy structure with version, statement, action, resource, and condition.

Explore how AWS IAM policies govern access by assigning users to groups, attaching read-only or administrator permissions, and crafting custom policies with JSON or a visual editor.

Learn to defend AWS accounts with strong password policies and multi-factor authentication. Explore virtual and physical MFA options, including Google Authenticator, Authy, YubiKey, Gemalto fobs, and GovCloud keys.

Configure the IAM password policy in the AWS console and enable MFA for the root account using an authenticator app.

Explore IAM roles that AWS services use to act on your behalf by granting permissions to resources such as EC2 instances, Lambda functions, and CloudFormation.

Practice creating an IAM role for an EC2 instance, attach the IAM read only access policy, set EC2 as the trusted service, and verify permissions in this hands on session.

Explore how encryption works in flight with tls/ssl and https, how server-side and client-side encryption protect data at rest and during transit, using data keys and tls certificates.

Explore AWS KMS, the service that manages encryption keys for AWS, enabling IAM access control, CloudTrail auditing, and seamless encryption across services like EBS, S3, and RDS.

Explore AWS KMS basics with AWS managed keys and customer managed keys, review key policies, cryptographic configuration, and perform encrypt and decrypt operations via CLI.

Discover and protect sensitive data in S3 with Macie, a fully managed service that uses machine learning and pattern matching to identify PII and alert via EventBridge.

AWS Secrets Manager enables rotating and generating secrets with AWS Lambda, integrates with RDS and Aurora databases, encrypts with KMS, and supports multi-region replication for disaster recovery.

Master AWS Secrets Manager to rotate, store, and retrieve secrets with database integrations and region replication. Learn encryption, IAM policies, and optional rotation via Lambda.

Explore how AWS WAF protects web apps at layer 7 with web ACLs, IP sets, and HTTP filters to block SQL injection and XSS.

Discover how AWS Shield protects against DDoS attacks, from Shield Standard’s free layer 3/4 protection to Shield Advanced’s automatic WAF rule deployment and 24/7 DDoS response.

Understand how a VPC and subnets in a region form private and public networks, with route tables guiding traffic between subnets and internet access via internet gateway and NAT gateway.

Explore network security in a VPC with network ACLs and security groups, and learn how VPC flow logs capture traffic for troubleshooting connectivity to S3, CloudWatch, or Kinesis.

Connect VPCs with peering, avoid overlapping IP ranges, and use VPC endpoints to access S3, DynamoDB, CloudWatch privately, plus site-to-site VPN or Direct Connect for on-premises connectivity.

Master core virtual private cloud concepts for AWS exam preparation. Identify default VPC, subnets, internet gateways, NAT gateways, NACLs, security groups, VPC peering, endpoints, flow logs, and on-premises connections.

Demonstrates how AWS PrivateLink creates private connectivity from a vendor's VPC to customer VPCs using a Network Load Balancer and Elastic Network Interface.